From 250ae2774923564daf0ccfa41eba997d64d98984 Mon Sep 17 00:00:00 2001 From: layerdiffusion <19834515+lllyasviel@users.noreply.github.com> Date: Tue, 20 Aug 2024 21:30:47 -0700 Subject: [PATCH] The First IDM-VTON that pass 4GB VRAM 100% reproduce official results --- .../forge_space_idm_vton/forge_app.py | 324 ++ .../forge_space_idm_vton/requirements.txt | 6 + .../forge_space_idm_vton/space_meta.json | 6 + .../src/attentionhacked_garmnet.py | 670 +++ .../src/attentionhacked_tryon.py | 682 ++++ .../src/transformerhacked_garmnet.py | 460 +++ .../src/transformerhacked_tryon.py | 467 +++ .../src/tryon_pipeline.py | 1893 +++++++++ .../src/unet_block_hacked_garmnet.py | 3579 +++++++++++++++++ .../src/unet_block_hacked_tryon.py | 3522 ++++++++++++++++ .../src/unet_hacked_garmnet.py | 1284 ++++++ .../src/unet_hacked_tryon.py | 1395 +++++++ 12 files changed, 14288 insertions(+) create mode 100644 extensions-builtin/forge_space_idm_vton/forge_app.py create mode 100644 extensions-builtin/forge_space_idm_vton/requirements.txt create mode 100644 extensions-builtin/forge_space_idm_vton/space_meta.json create mode 100644 extensions-builtin/forge_space_idm_vton/src/attentionhacked_garmnet.py create mode 100644 extensions-builtin/forge_space_idm_vton/src/attentionhacked_tryon.py create mode 100644 extensions-builtin/forge_space_idm_vton/src/transformerhacked_garmnet.py create mode 100644 extensions-builtin/forge_space_idm_vton/src/transformerhacked_tryon.py create mode 100644 extensions-builtin/forge_space_idm_vton/src/tryon_pipeline.py create mode 100644 extensions-builtin/forge_space_idm_vton/src/unet_block_hacked_garmnet.py create mode 100644 extensions-builtin/forge_space_idm_vton/src/unet_block_hacked_tryon.py create mode 100644 extensions-builtin/forge_space_idm_vton/src/unet_hacked_garmnet.py create mode 100644 extensions-builtin/forge_space_idm_vton/src/unet_hacked_tryon.py diff --git a/extensions-builtin/forge_space_idm_vton/forge_app.py b/extensions-builtin/forge_space_idm_vton/forge_app.py new file mode 100644 index 00000000..1d982a82 --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/forge_app.py @@ -0,0 +1,324 @@ +import spaces +import contextlib + +import gradio as gr +from PIL import Image +from src.tryon_pipeline import StableDiffusionXLInpaintPipeline as TryonPipeline +from src.unet_hacked_garmnet import UNet2DConditionModel as UNet2DConditionModel_ref +from src.unet_hacked_tryon import UNet2DConditionModel +from transformers import ( + CLIPImageProcessor, + CLIPVisionModelWithProjection, + CLIPTextModel, + CLIPTextModelWithProjection, +) +from diffusers import DDPMScheduler, AutoencoderKL +from typing import List + +import torch +import os +from transformers import AutoTokenizer +import numpy as np +from utils_mask import get_mask_location +from torchvision import transforms +import apply_net +from preprocess.humanparsing.run_parsing import Parsing +from preprocess.openpose.run_openpose import OpenPose +from detectron2.data.detection_utils import convert_PIL_to_numpy, _apply_exif_orientation +from torchvision.transforms.functional import to_pil_image + + +def pil_to_binary_mask(pil_image, threshold=0): + np_image = np.array(pil_image) + grayscale_image = Image.fromarray(np_image).convert("L") + binary_mask = np.array(grayscale_image) > threshold + mask = np.zeros(binary_mask.shape, dtype=np.uint8) + for i in range(binary_mask.shape[0]): + for j in range(binary_mask.shape[1]): + if binary_mask[i, j] == True: + mask[i, j] = 1 + mask = (mask * 255).astype(np.uint8) + output_mask = Image.fromarray(mask) + return output_mask + + +base_path = 'yisol/IDM-VTON' +example_path = os.path.join(spaces.convert_root_path(), 'example') + + +with spaces.capture_gpu_object() as GO: + unet = UNet2DConditionModel.from_pretrained( + base_path, + subfolder="unet", + torch_dtype=torch.float16, + ) + unet.requires_grad_(False) + tokenizer_one = AutoTokenizer.from_pretrained( + base_path, + subfolder="tokenizer", + revision=None, + use_fast=False, + ) + tokenizer_two = AutoTokenizer.from_pretrained( + base_path, + subfolder="tokenizer_2", + revision=None, + use_fast=False, + ) + noise_scheduler = DDPMScheduler.from_pretrained(base_path, subfolder="scheduler") + + text_encoder_one = CLIPTextModel.from_pretrained( + base_path, + subfolder="text_encoder", + torch_dtype=torch.float16, + ) + text_encoder_two = CLIPTextModelWithProjection.from_pretrained( + base_path, + subfolder="text_encoder_2", + torch_dtype=torch.float16, + ) + image_encoder = CLIPVisionModelWithProjection.from_pretrained( + base_path, + subfolder="image_encoder", + torch_dtype=torch.float16, + ) + vae = AutoencoderKL.from_pretrained(base_path, + subfolder="vae", + torch_dtype=torch.float16, + ) + + # "stabilityai/stable-diffusion-xl-base-1.0", + UNet_Encoder = UNet2DConditionModel_ref.from_pretrained( + base_path, + subfolder="unet_encoder", + torch_dtype=torch.float16, + ) + + parsing_model = Parsing(0) + openpose_model = OpenPose(0) + + UNet_Encoder.requires_grad_(False) + image_encoder.requires_grad_(False) + vae.requires_grad_(False) + unet.requires_grad_(False) + text_encoder_one.requires_grad_(False) + text_encoder_two.requires_grad_(False) + tensor_transfrom = transforms.Compose( + [ + transforms.ToTensor(), + transforms.Normalize([0.5], [0.5]), + ] + ) + + pipe = TryonPipeline.from_pretrained( + base_path, + unet=unet, + vae=vae, + feature_extractor=CLIPImageProcessor(), + text_encoder=text_encoder_one, + text_encoder_2=text_encoder_two, + tokenizer=tokenizer_one, + tokenizer_2=tokenizer_two, + scheduler=noise_scheduler, + image_encoder=image_encoder, + torch_dtype=torch.float16, + ) + pipe.unet_encoder = UNet_Encoder + +unet_joint = torch.nn.ModuleList([ + unet, UNet_Encoder +]) + + +spaces.automatically_move_to_gpu_when_forward(unet, target_model=unet_joint) +spaces.automatically_move_to_gpu_when_forward(UNet_Encoder, target_model=unet_joint) +spaces.automatically_move_to_gpu_when_forward(unet.encoder_hid_proj) + +spaces.automatically_move_pipeline_components(pipe=pipe) + +spaces.automatically_move_to_gpu_when_forward(openpose_model.preprocessor.body_estimation.model) +spaces.change_attention_from_diffusers_to_forge(unet) +spaces.change_attention_from_diffusers_to_forge(vae) +spaces.change_attention_from_diffusers_to_forge(UNet_Encoder) + + +@spaces.GPU(gpu_objects=GO, manual_load=True) +def start_tryon(dict, garm_img, garment_des, is_checked, is_checked_crop, denoise_steps, seed, progress=gr.Progress(track_tqdm=True)): + device = spaces.gpu + + # openpose_model.preprocessor.body_estimation.model.to(device) + # pipe.to(device) + # pipe.unet_encoder.to(device) + + garm_img = garm_img.convert("RGB").resize((768, 1024)) + human_img_orig = dict["background"].convert("RGB") + + if is_checked_crop: + width, height = human_img_orig.size + target_width = int(min(width, height * (3 / 4))) + target_height = int(min(height, width * (4 / 3))) + left = (width - target_width) / 2 + top = (height - target_height) / 2 + right = (width + target_width) / 2 + bottom = (height + target_height) / 2 + cropped_img = human_img_orig.crop((left, top, right, bottom)) + crop_size = cropped_img.size + human_img = cropped_img.resize((768, 1024)) + else: + human_img = human_img_orig.resize((768, 1024)) + + if is_checked: + keypoints = openpose_model(human_img.resize((384, 512))) + model_parse, _ = parsing_model(human_img.resize((384, 512))) + mask, mask_gray = get_mask_location('hd', "upper_body", model_parse, keypoints) + mask = mask.resize((768, 1024)) + else: + mask = pil_to_binary_mask(dict['layers'][0].convert("RGB").resize((768, 1024))) + # mask = transforms.ToTensor()(mask) + # mask = mask.unsqueeze(0) + mask_gray = (1 - transforms.ToTensor()(mask)) * tensor_transfrom(human_img) + mask_gray = to_pil_image((mask_gray + 1.0) / 2.0) + + human_img_arg = _apply_exif_orientation(human_img.resize((384, 512))) + human_img_arg = convert_PIL_to_numpy(human_img_arg, format="BGR") + + args = apply_net.create_argument_parser().parse_args(('show', spaces.convert_root_path() + 'configs/densepose_rcnn_R_50_FPN_s1x.yaml', spaces.convert_root_path() + 'ckpt/densepose/model_final_162be9.pkl', 'dp_segm', '-v', '--opts', 'MODEL.DEVICE', 'cuda')) + # verbosity = getattr(args, "verbosity", None) + pose_img = args.func(args, human_img_arg) + pose_img = pose_img[:, :, ::-1] + pose_img = Image.fromarray(pose_img).resize((768, 1024)) + + with torch.no_grad(): + # Extract the images + with torch.cuda.amp.autocast(): + with torch.no_grad(): + prompt = "model is wearing " + garment_des + negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality" + with torch.inference_mode(): + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = pipe.encode_prompt( + prompt, + num_images_per_prompt=1, + do_classifier_free_guidance=True, + negative_prompt=negative_prompt, + ) + + prompt = "a photo of " + garment_des + negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality" + if not isinstance(prompt, List): + prompt = [prompt] * 1 + if not isinstance(negative_prompt, List): + negative_prompt = [negative_prompt] * 1 + with torch.inference_mode(): + ( + prompt_embeds_c, + _, + _, + _, + ) = pipe.encode_prompt( + prompt, + num_images_per_prompt=1, + do_classifier_free_guidance=False, + negative_prompt=negative_prompt, + ) + + pose_img = tensor_transfrom(pose_img).unsqueeze(0).to(device, torch.float16) + garm_tensor = tensor_transfrom(garm_img).unsqueeze(0).to(device, torch.float16) + generator = torch.Generator(device).manual_seed(seed) if seed is not None else None + images = pipe( + prompt_embeds=prompt_embeds.to(device, torch.float16), + negative_prompt_embeds=negative_prompt_embeds.to(device, torch.float16), + pooled_prompt_embeds=pooled_prompt_embeds.to(device, torch.float16), + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds.to(device, torch.float16), + num_inference_steps=denoise_steps, + generator=generator, + strength=1.0, + pose_img=pose_img.to(device, torch.float16), + text_embeds_cloth=prompt_embeds_c.to(device, torch.float16), + cloth=garm_tensor.to(device, torch.float16), + mask_image=mask, + image=human_img, + height=1024, + width=768, + ip_adapter_image=garm_img.resize((768, 1024)), + guidance_scale=2.0, + )[0] + + if is_checked_crop: + out_img = images[0].resize(crop_size) + human_img_orig.paste(out_img, (int(left), int(top))) + return human_img_orig, mask_gray + else: + return images[0], mask_gray + # return images[0], mask_gray + + +garm_list = os.listdir(os.path.join(example_path, "cloth")) +garm_list_path = [os.path.join(example_path, "cloth", garm) for garm in garm_list] + +human_list = os.listdir(os.path.join(example_path, "human")) +human_list_path = [os.path.join(example_path, "human", human) for human in human_list] + +human_ex_list = [] +for ex_human in human_list_path: + ex_dict = {} + ex_dict['background'] = ex_human + ex_dict['layers'] = None + ex_dict['composite'] = None + human_ex_list.append(ex_dict) + +##default human + + +image_blocks = gr.Blocks().queue() +with image_blocks as demo: + gr.Markdown("## IDM-VTON ๐Ÿ‘•๐Ÿ‘”๐Ÿ‘š") + gr.Markdown("Virtual Try-on with your image and garment image. Check out the [source codes](https://github.com/yisol/IDM-VTON) and the [model](https://huggingface.co/yisol/IDM-VTON)") + with gr.Row(): + with gr.Column(): + imgs = gr.ImageEditor(sources='upload', type="pil", label='Human. Mask with pen or use auto-masking', interactive=True) + with gr.Row(): + is_checked = gr.Checkbox(label="Yes", info="Use auto-generated mask (Takes 5 seconds)", value=True) + with gr.Row(): + is_checked_crop = gr.Checkbox(label="Yes", info="Use auto-crop & resizing", value=False) + + example = gr.Examples( + inputs=imgs, + examples_per_page=10, + examples=human_ex_list + ) + + with gr.Column(): + garm_img = gr.Image(label="Garment", sources='upload', type="pil") + with gr.Row(elem_id="prompt-container"): + with gr.Row(): + prompt = gr.Textbox(placeholder="Description of garment ex) Short Sleeve Round Neck T-shirts", show_label=False, elem_id="prompt") + example = gr.Examples( + inputs=garm_img, + examples_per_page=8, + examples=garm_list_path) + with gr.Column(): + # image_out = gr.Image(label="Output", elem_id="output-img", height=400) + masked_img = gr.Image(label="Masked image output", elem_id="masked-img", show_share_button=False) + with gr.Column(): + # image_out = gr.Image(label="Output", elem_id="output-img", height=400) + image_out = gr.Image(label="Output", elem_id="output-img", show_share_button=False) + + with gr.Column(): + try_button = gr.Button(value="Try-on") + with gr.Accordion(label="Advanced Settings", open=False): + with gr.Row(): + denoise_steps = gr.Number(label="Denoising Steps", minimum=20, maximum=40, value=30, step=1) + seed = gr.Number(label="Seed", minimum=-1, maximum=2147483647, step=1, value=42) + + try_button.click(fn=start_tryon, inputs=[imgs, garm_img, prompt, is_checked, is_checked_crop, denoise_steps, seed], outputs=[image_out, masked_img], api_name='tryon') + + +demo = image_blocks + +if __name__ == "__main__": + demo.launch() diff --git a/extensions-builtin/forge_space_idm_vton/requirements.txt b/extensions-builtin/forge_space_idm_vton/requirements.txt new file mode 100644 index 00000000..a649adfb --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/requirements.txt @@ -0,0 +1,6 @@ +basicsr +fvcore +cloudpickle +omegaconf +pycocotools +config==0.5.1 diff --git a/extensions-builtin/forge_space_idm_vton/space_meta.json b/extensions-builtin/forge_space_idm_vton/space_meta.json new file mode 100644 index 00000000..172fff3b --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/space_meta.json @@ -0,0 +1,6 @@ +{ + "tag": "General Image Processing and Applications", + "title": "IDM-VTON: Virtual Try-on with your image and garment image", + "repo_id": "yisol/IDM-VTON", + "revision": "810e5908d53e4023f3dade2b8ebf10e3ae995be1" +} diff --git a/extensions-builtin/forge_space_idm_vton/src/attentionhacked_garmnet.py b/extensions-builtin/forge_space_idm_vton/src/attentionhacked_garmnet.py new file mode 100644 index 00000000..66885c47 --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/src/attentionhacked_garmnet.py @@ -0,0 +1,670 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Any, Dict, Optional + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND +from diffusers.utils.torch_utils import maybe_allow_in_graph +from diffusers.models.activations import GEGLU, GELU, ApproximateGELU +from diffusers.models.attention_processor import Attention +from diffusers.models.embeddings import SinusoidalPositionalEmbedding +from diffusers.models.lora import LoRACompatibleLinear +from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm + + +def _chunked_feed_forward( + ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None +): + # "feed_forward_chunk_size" can be used to save memory + if hidden_states.shape[chunk_dim] % chunk_size != 0: + raise ValueError( + f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`." + ) + + num_chunks = hidden_states.shape[chunk_dim] // chunk_size + if lora_scale is None: + ff_output = torch.cat( + [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)], + dim=chunk_dim, + ) + else: + # TOOD(Patrick): LoRA scale can be removed once PEFT refactor is complete + ff_output = torch.cat( + [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)], + dim=chunk_dim, + ) + + return ff_output + + +@maybe_allow_in_graph +class GatedSelfAttentionDense(nn.Module): + r""" + A gated self-attention dense layer that combines visual features and object features. + + Parameters: + query_dim (`int`): The number of channels in the query. + context_dim (`int`): The number of channels in the context. + n_heads (`int`): The number of heads to use for attention. + d_head (`int`): The number of channels in each head. + """ + + def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int): + super().__init__() + + # we need a linear projection since we need cat visual feature and obj feature + self.linear = nn.Linear(context_dim, query_dim) + + self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head) + self.ff = FeedForward(query_dim, activation_fn="geglu") + + self.norm1 = nn.LayerNorm(query_dim) + self.norm2 = nn.LayerNorm(query_dim) + + self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0))) + self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0))) + + self.enabled = True + + def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor: + if not self.enabled: + return x + + n_visual = x.shape[1] + objs = self.linear(objs) + + x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :] + x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x)) + + return x + + +@maybe_allow_in_graph +class BasicTransformerBlock(nn.Module): + r""" + A basic Transformer block. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm (: + obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`. + attention_bias (: + obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter. + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used. + double_self_attention (`bool`, *optional*): + Whether to use two self-attention layers. In this case no cross attention layers are used. + upcast_attention (`bool`, *optional*): + Whether to upcast the attention computation to float32. This is useful for mixed precision training. + norm_elementwise_affine (`bool`, *optional*, defaults to `True`): + Whether to use learnable elementwise affine parameters for normalization. + norm_type (`str`, *optional*, defaults to `"layer_norm"`): + The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`. + final_dropout (`bool` *optional*, defaults to False): + Whether to apply a final dropout after the last feed-forward layer. + attention_type (`str`, *optional*, defaults to `"default"`): + The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`. + positional_embeddings (`str`, *optional*, defaults to `None`): + The type of positional embeddings to apply to. + num_positional_embeddings (`int`, *optional*, defaults to `None`): + The maximum number of positional embeddings to apply. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + attention_bias: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_elementwise_affine: bool = True, + norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single' + norm_eps: float = 1e-5, + final_dropout: bool = False, + attention_type: str = "default", + positional_embeddings: Optional[str] = None, + num_positional_embeddings: Optional[int] = None, + ada_norm_continous_conditioning_embedding_dim: Optional[int] = None, + ada_norm_bias: Optional[int] = None, + ff_inner_dim: Optional[int] = None, + ff_bias: bool = True, + attention_out_bias: bool = True, + ): + super().__init__() + self.only_cross_attention = only_cross_attention + + self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero" + self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" + self.use_ada_layer_norm_single = norm_type == "ada_norm_single" + self.use_layer_norm = norm_type == "layer_norm" + self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous" + + if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: + raise ValueError( + f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" + f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." + ) + + if positional_embeddings and (num_positional_embeddings is None): + raise ValueError( + "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined." + ) + + if positional_embeddings == "sinusoidal": + self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings) + else: + self.pos_embed = None + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + if self.use_ada_layer_norm: + self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif self.use_ada_layer_norm_zero: + self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm) + elif self.use_ada_layer_norm_continuous: + self.norm1 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim if only_cross_attention else None, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None or double_self_attention: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + if self.use_ada_layer_norm: + self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif self.use_ada_layer_norm_continuous: + self.norm2 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim if not double_self_attention else None, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) # is self-attn if encoder_hidden_states is none + else: + self.norm2 = None + self.attn2 = None + + # 3. Feed-forward + if self.use_ada_layer_norm_continuous: + self.norm3 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "layer_norm", + ) + elif not self.use_ada_layer_norm_single: + self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + + self.ff = FeedForward( + dim, + dropout=dropout, + activation_fn=activation_fn, + final_dropout=final_dropout, + inner_dim=ff_inner_dim, + bias=ff_bias, + ) + + # 4. Fuser + if attention_type == "gated" or attention_type == "gated-text-image": + self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim) + + # 5. Scale-shift for PixArt-Alpha. + if self.use_ada_layer_norm_single: + self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5) + + # let chunk size default to None + self._chunk_size = None + self._chunk_dim = 0 + + def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0): + # Sets chunk feed-forward + self._chunk_size = chunk_size + self._chunk_dim = dim + + def forward( + self, + hidden_states: torch.FloatTensor, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + timestep: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + class_labels: Optional[torch.LongTensor] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + ) -> torch.FloatTensor: + # Notice that normalization is always applied before the real computation in the following blocks. + # 0. Self-Attention + batch_size = hidden_states.shape[0] + if self.use_ada_layer_norm: + norm_hidden_states = self.norm1(hidden_states, timestep) + elif self.use_ada_layer_norm_zero: + norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1( + hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + elif self.use_layer_norm: + norm_hidden_states = self.norm1(hidden_states) + elif self.use_ada_layer_norm_continuous: + norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif self.use_ada_layer_norm_single: + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1) + ).chunk(6, dim=1) + norm_hidden_states = self.norm1(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa + norm_hidden_states = norm_hidden_states.squeeze(1) + else: + raise ValueError("Incorrect norm used") + + if self.pos_embed is not None: + norm_hidden_states = self.pos_embed(norm_hidden_states) + + garment_features = [] + garment_features.append(norm_hidden_states) + + # 1. Retrieve lora scale. + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + # 2. Prepare GLIGEN inputs + cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} + gligen_kwargs = cross_attention_kwargs.pop("gligen", None) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + if self.use_ada_layer_norm_zero: + attn_output = gate_msa.unsqueeze(1) * attn_output + elif self.use_ada_layer_norm_single: + attn_output = gate_msa * attn_output + + hidden_states = attn_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + # 2.5 GLIGEN Control + if gligen_kwargs is not None: + hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"]) + + # 3. Cross-Attention + if self.attn2 is not None: + if self.use_ada_layer_norm: + norm_hidden_states = self.norm2(hidden_states, timestep) + elif self.use_ada_layer_norm_zero or self.use_layer_norm: + norm_hidden_states = self.norm2(hidden_states) + elif self.use_ada_layer_norm_single: + # For PixArt norm2 isn't applied here: + # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103 + norm_hidden_states = hidden_states + elif self.use_ada_layer_norm_continuous: + norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"]) + else: + raise ValueError("Incorrect norm") + + if self.pos_embed is not None and self.use_ada_layer_norm_single is False: + norm_hidden_states = self.pos_embed(norm_hidden_states) + + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + hidden_states = attn_output + hidden_states + + # 4. Feed-forward + if self.use_ada_layer_norm_continuous: + norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif not self.use_ada_layer_norm_single: + norm_hidden_states = self.norm3(hidden_states) + + if self.use_ada_layer_norm_zero: + norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] + + if self.use_ada_layer_norm_single: + norm_hidden_states = self.norm2(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp + + if self._chunk_size is not None: + # "feed_forward_chunk_size" can be used to save memory + ff_output = _chunked_feed_forward( + self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale + ) + else: + ff_output = self.ff(norm_hidden_states, scale=lora_scale) + + if self.use_ada_layer_norm_zero: + ff_output = gate_mlp.unsqueeze(1) * ff_output + elif self.use_ada_layer_norm_single: + ff_output = gate_mlp * ff_output + + hidden_states = ff_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + return hidden_states, garment_features + + +@maybe_allow_in_graph +class TemporalBasicTransformerBlock(nn.Module): + r""" + A basic Transformer block for video like data. + + Parameters: + dim (`int`): The number of channels in the input and output. + time_mix_inner_dim (`int`): The number of channels for temporal attention. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + """ + + def __init__( + self, + dim: int, + time_mix_inner_dim: int, + num_attention_heads: int, + attention_head_dim: int, + cross_attention_dim: Optional[int] = None, + ): + super().__init__() + self.is_res = dim == time_mix_inner_dim + + self.norm_in = nn.LayerNorm(dim) + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + self.norm_in = nn.LayerNorm(dim) + self.ff_in = FeedForward( + dim, + dim_out=time_mix_inner_dim, + activation_fn="geglu", + ) + + self.norm1 = nn.LayerNorm(time_mix_inner_dim) + self.attn1 = Attention( + query_dim=time_mix_inner_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + cross_attention_dim=None, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + self.norm2 = nn.LayerNorm(time_mix_inner_dim) + self.attn2 = Attention( + query_dim=time_mix_inner_dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + ) # is self-attn if encoder_hidden_states is none + else: + self.norm2 = None + self.attn2 = None + + # 3. Feed-forward + self.norm3 = nn.LayerNorm(time_mix_inner_dim) + self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu") + + # let chunk size default to None + self._chunk_size = None + self._chunk_dim = None + + def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs): + # Sets chunk feed-forward + self._chunk_size = chunk_size + # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off + self._chunk_dim = 1 + + def forward( + self, + hidden_states: torch.FloatTensor, + num_frames: int, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + # Notice that normalization is always applied before the real computation in the following blocks. + # 0. Self-Attention + batch_size = hidden_states.shape[0] + + batch_frames, seq_length, channels = hidden_states.shape + batch_size = batch_frames // num_frames + + hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels) + hidden_states = hidden_states.permute(0, 2, 1, 3) + hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels) + + residual = hidden_states + hidden_states = self.norm_in(hidden_states) + + if self._chunk_size is not None: + hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size) + else: + hidden_states = self.ff_in(hidden_states) + + if self.is_res: + hidden_states = hidden_states + residual + + norm_hidden_states = self.norm1(hidden_states) + attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None) + hidden_states = attn_output + hidden_states + + # 3. Cross-Attention + if self.attn2 is not None: + norm_hidden_states = self.norm2(hidden_states) + attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states) + hidden_states = attn_output + hidden_states + + # 4. Feed-forward + norm_hidden_states = self.norm3(hidden_states) + + if self._chunk_size is not None: + ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size) + else: + ff_output = self.ff(norm_hidden_states) + + if self.is_res: + hidden_states = ff_output + hidden_states + else: + hidden_states = ff_output + + hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels) + hidden_states = hidden_states.permute(0, 2, 1, 3) + hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels) + + return hidden_states + + +class SkipFFTransformerBlock(nn.Module): + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + kv_input_dim: int, + kv_input_dim_proj_use_bias: bool, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + attention_out_bias: bool = True, + ): + super().__init__() + if kv_input_dim != dim: + self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias) + else: + self.kv_mapper = None + + self.norm1 = RMSNorm(dim, 1e-06) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim, + out_bias=attention_out_bias, + ) + + self.norm2 = RMSNorm(dim, 1e-06) + + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + out_bias=attention_out_bias, + ) + + def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs): + cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} + + if self.kv_mapper is not None: + encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states)) + + norm_hidden_states = self.norm1(hidden_states) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + norm_hidden_states = self.norm2(hidden_states) + + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + return hidden_states + + +class FeedForward(nn.Module): + r""" + A feed-forward layer. + + Parameters: + dim (`int`): The number of channels in the input. + dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`. + mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + final_dropout (`bool` *optional*, defaults to False): Apply a final dropout. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__( + self, + dim: int, + dim_out: Optional[int] = None, + mult: int = 4, + dropout: float = 0.0, + activation_fn: str = "geglu", + final_dropout: bool = False, + inner_dim=None, + bias: bool = True, + ): + super().__init__() + if inner_dim is None: + inner_dim = int(dim * mult) + dim_out = dim_out if dim_out is not None else dim + linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear + + if activation_fn == "gelu": + act_fn = GELU(dim, inner_dim, bias=bias) + if activation_fn == "gelu-approximate": + act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias) + elif activation_fn == "geglu": + act_fn = GEGLU(dim, inner_dim, bias=bias) + elif activation_fn == "geglu-approximate": + act_fn = ApproximateGELU(dim, inner_dim, bias=bias) + + self.net = nn.ModuleList([]) + # project in + self.net.append(act_fn) + # project dropout + self.net.append(nn.Dropout(dropout)) + # project out + self.net.append(linear_cls(inner_dim, dim_out, bias=bias)) + # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout + if final_dropout: + self.net.append(nn.Dropout(dropout)) + + def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor: + compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear) + for module in self.net: + if isinstance(module, compatible_cls): + hidden_states = module(hidden_states, scale) + else: + hidden_states = module(hidden_states) + return hidden_states diff --git a/extensions-builtin/forge_space_idm_vton/src/attentionhacked_tryon.py b/extensions-builtin/forge_space_idm_vton/src/attentionhacked_tryon.py new file mode 100644 index 00000000..9947e7c8 --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/src/attentionhacked_tryon.py @@ -0,0 +1,682 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Any, Dict, Optional + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND +from diffusers.utils.torch_utils import maybe_allow_in_graph +from diffusers.models.activations import GEGLU, GELU, ApproximateGELU +from diffusers.models.attention_processor import Attention +from diffusers.models.embeddings import SinusoidalPositionalEmbedding +from diffusers.models.lora import LoRACompatibleLinear +from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm + + +def _chunked_feed_forward( + ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None +): + # "feed_forward_chunk_size" can be used to save memory + if hidden_states.shape[chunk_dim] % chunk_size != 0: + raise ValueError( + f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`." + ) + + num_chunks = hidden_states.shape[chunk_dim] // chunk_size + if lora_scale is None: + ff_output = torch.cat( + [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)], + dim=chunk_dim, + ) + else: + # TOOD(Patrick): LoRA scale can be removed once PEFT refactor is complete + ff_output = torch.cat( + [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)], + dim=chunk_dim, + ) + + return ff_output + + +@maybe_allow_in_graph +class GatedSelfAttentionDense(nn.Module): + r""" + A gated self-attention dense layer that combines visual features and object features. + + Parameters: + query_dim (`int`): The number of channels in the query. + context_dim (`int`): The number of channels in the context. + n_heads (`int`): The number of heads to use for attention. + d_head (`int`): The number of channels in each head. + """ + + def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int): + super().__init__() + + # we need a linear projection since we need cat visual feature and obj feature + self.linear = nn.Linear(context_dim, query_dim) + + self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head) + self.ff = FeedForward(query_dim, activation_fn="geglu") + + self.norm1 = nn.LayerNorm(query_dim) + self.norm2 = nn.LayerNorm(query_dim) + + self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0))) + self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0))) + + self.enabled = True + + def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor: + if not self.enabled: + return x + + n_visual = x.shape[1] + objs = self.linear(objs) + + x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :] + x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x)) + + return x + + +@maybe_allow_in_graph +class BasicTransformerBlock(nn.Module): + r""" + A basic Transformer block. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm (: + obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`. + attention_bias (: + obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter. + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used. + double_self_attention (`bool`, *optional*): + Whether to use two self-attention layers. In this case no cross attention layers are used. + upcast_attention (`bool`, *optional*): + Whether to upcast the attention computation to float32. This is useful for mixed precision training. + norm_elementwise_affine (`bool`, *optional*, defaults to `True`): + Whether to use learnable elementwise affine parameters for normalization. + norm_type (`str`, *optional*, defaults to `"layer_norm"`): + The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`. + final_dropout (`bool` *optional*, defaults to False): + Whether to apply a final dropout after the last feed-forward layer. + attention_type (`str`, *optional*, defaults to `"default"`): + The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`. + positional_embeddings (`str`, *optional*, defaults to `None`): + The type of positional embeddings to apply to. + num_positional_embeddings (`int`, *optional*, defaults to `None`): + The maximum number of positional embeddings to apply. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + attention_bias: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_elementwise_affine: bool = True, + norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single' + norm_eps: float = 1e-5, + final_dropout: bool = False, + attention_type: str = "default", + positional_embeddings: Optional[str] = None, + num_positional_embeddings: Optional[int] = None, + ada_norm_continous_conditioning_embedding_dim: Optional[int] = None, + ada_norm_bias: Optional[int] = None, + ff_inner_dim: Optional[int] = None, + ff_bias: bool = True, + attention_out_bias: bool = True, + ): + super().__init__() + self.only_cross_attention = only_cross_attention + + self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero" + self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" + self.use_ada_layer_norm_single = norm_type == "ada_norm_single" + self.use_layer_norm = norm_type == "layer_norm" + self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous" + + if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: + raise ValueError( + f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" + f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." + ) + + if positional_embeddings and (num_positional_embeddings is None): + raise ValueError( + "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined." + ) + + if positional_embeddings == "sinusoidal": + self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings) + else: + self.pos_embed = None + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + if self.use_ada_layer_norm: + self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif self.use_ada_layer_norm_zero: + self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm) + elif self.use_ada_layer_norm_continuous: + self.norm1 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim if only_cross_attention else None, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None or double_self_attention: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + if self.use_ada_layer_norm: + self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif self.use_ada_layer_norm_continuous: + self.norm2 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim if not double_self_attention else None, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) # is self-attn if encoder_hidden_states is none + else: + self.norm2 = None + self.attn2 = None + + # 3. Feed-forward + if self.use_ada_layer_norm_continuous: + self.norm3 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "layer_norm", + ) + elif not self.use_ada_layer_norm_single: + self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + + self.ff = FeedForward( + dim, + dropout=dropout, + activation_fn=activation_fn, + final_dropout=final_dropout, + inner_dim=ff_inner_dim, + bias=ff_bias, + ) + + # 4. Fuser + if attention_type == "gated" or attention_type == "gated-text-image": + self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim) + + # 5. Scale-shift for PixArt-Alpha. + if self.use_ada_layer_norm_single: + self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5) + + # let chunk size default to None + self._chunk_size = None + self._chunk_dim = 0 + + def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0): + # Sets chunk feed-forward + self._chunk_size = chunk_size + self._chunk_dim = dim + + def forward( + self, + hidden_states: torch.FloatTensor, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + timestep: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + class_labels: Optional[torch.LongTensor] = None, + garment_features=None, + curr_garment_feat_idx=0, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + ) -> torch.FloatTensor: + # Notice that normalization is always applied before the real computation in the following blocks. + # 0. Self-Attention + batch_size = hidden_states.shape[0] + + + + if self.use_ada_layer_norm: + norm_hidden_states = self.norm1(hidden_states, timestep) + elif self.use_ada_layer_norm_zero: + norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1( + hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + elif self.use_layer_norm: + norm_hidden_states = self.norm1(hidden_states) + elif self.use_ada_layer_norm_continuous: + norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif self.use_ada_layer_norm_single: + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1) + ).chunk(6, dim=1) + norm_hidden_states = self.norm1(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa + norm_hidden_states = norm_hidden_states.squeeze(1) + else: + raise ValueError("Incorrect norm used") + + if self.pos_embed is not None: + norm_hidden_states = self.pos_embed(norm_hidden_states) + + # 1. Retrieve lora scale. + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + # 2. Prepare GLIGEN inputs + cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} + gligen_kwargs = cross_attention_kwargs.pop("gligen", None) + + + #type2 + modify_norm_hidden_states = torch.cat([norm_hidden_states,garment_features[curr_garment_feat_idx]], dim=1) + curr_garment_feat_idx +=1 + attn_output = self.attn1( + #norm_hidden_states, + modify_norm_hidden_states, + encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + if self.use_ada_layer_norm_zero: + attn_output = gate_msa.unsqueeze(1) * attn_output + elif self.use_ada_layer_norm_single: + attn_output = gate_msa * attn_output + + + #type2 + hidden_states = attn_output[:,:hidden_states.shape[-2],:] + hidden_states + + + + + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + # 2.5 GLIGEN Control + if gligen_kwargs is not None: + hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"]) + + # 3. Cross-Attention + if self.attn2 is not None: + if self.use_ada_layer_norm: + norm_hidden_states = self.norm2(hidden_states, timestep) + elif self.use_ada_layer_norm_zero or self.use_layer_norm: + norm_hidden_states = self.norm2(hidden_states) + elif self.use_ada_layer_norm_single: + # For PixArt norm2 isn't applied here: + # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103 + norm_hidden_states = hidden_states + elif self.use_ada_layer_norm_continuous: + norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"]) + else: + raise ValueError("Incorrect norm") + + if self.pos_embed is not None and self.use_ada_layer_norm_single is False: + norm_hidden_states = self.pos_embed(norm_hidden_states) + + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + hidden_states = attn_output + hidden_states + + # 4. Feed-forward + if self.use_ada_layer_norm_continuous: + norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif not self.use_ada_layer_norm_single: + norm_hidden_states = self.norm3(hidden_states) + + if self.use_ada_layer_norm_zero: + norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] + + if self.use_ada_layer_norm_single: + norm_hidden_states = self.norm2(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp + + if self._chunk_size is not None: + # "feed_forward_chunk_size" can be used to save memory + ff_output = _chunked_feed_forward( + self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale + ) + else: + ff_output = self.ff(norm_hidden_states, scale=lora_scale) + + if self.use_ada_layer_norm_zero: + ff_output = gate_mlp.unsqueeze(1) * ff_output + elif self.use_ada_layer_norm_single: + ff_output = gate_mlp * ff_output + + hidden_states = ff_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + return hidden_states,curr_garment_feat_idx + + +@maybe_allow_in_graph +class TemporalBasicTransformerBlock(nn.Module): + r""" + A basic Transformer block for video like data. + + Parameters: + dim (`int`): The number of channels in the input and output. + time_mix_inner_dim (`int`): The number of channels for temporal attention. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + """ + + def __init__( + self, + dim: int, + time_mix_inner_dim: int, + num_attention_heads: int, + attention_head_dim: int, + cross_attention_dim: Optional[int] = None, + ): + super().__init__() + self.is_res = dim == time_mix_inner_dim + + self.norm_in = nn.LayerNorm(dim) + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + self.norm_in = nn.LayerNorm(dim) + self.ff_in = FeedForward( + dim, + dim_out=time_mix_inner_dim, + activation_fn="geglu", + ) + + self.norm1 = nn.LayerNorm(time_mix_inner_dim) + self.attn1 = Attention( + query_dim=time_mix_inner_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + cross_attention_dim=None, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + self.norm2 = nn.LayerNorm(time_mix_inner_dim) + self.attn2 = Attention( + query_dim=time_mix_inner_dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + ) # is self-attn if encoder_hidden_states is none + else: + self.norm2 = None + self.attn2 = None + + # 3. Feed-forward + self.norm3 = nn.LayerNorm(time_mix_inner_dim) + self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu") + + # let chunk size default to None + self._chunk_size = None + self._chunk_dim = None + + def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs): + # Sets chunk feed-forward + self._chunk_size = chunk_size + # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off + self._chunk_dim = 1 + + def forward( + self, + hidden_states: torch.FloatTensor, + num_frames: int, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + # Notice that normalization is always applied before the real computation in the following blocks. + # 0. Self-Attention + batch_size = hidden_states.shape[0] + + batch_frames, seq_length, channels = hidden_states.shape + batch_size = batch_frames // num_frames + + hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels) + hidden_states = hidden_states.permute(0, 2, 1, 3) + hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels) + + residual = hidden_states + hidden_states = self.norm_in(hidden_states) + + if self._chunk_size is not None: + hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size) + else: + hidden_states = self.ff_in(hidden_states) + + if self.is_res: + hidden_states = hidden_states + residual + + norm_hidden_states = self.norm1(hidden_states) + attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None) + hidden_states = attn_output + hidden_states + + # 3. Cross-Attention + if self.attn2 is not None: + norm_hidden_states = self.norm2(hidden_states) + attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states) + hidden_states = attn_output + hidden_states + + # 4. Feed-forward + norm_hidden_states = self.norm3(hidden_states) + + if self._chunk_size is not None: + ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size) + else: + ff_output = self.ff(norm_hidden_states) + + if self.is_res: + hidden_states = ff_output + hidden_states + else: + hidden_states = ff_output + + hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels) + hidden_states = hidden_states.permute(0, 2, 1, 3) + hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels) + + return hidden_states + + +class SkipFFTransformerBlock(nn.Module): + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + kv_input_dim: int, + kv_input_dim_proj_use_bias: bool, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + attention_out_bias: bool = True, + ): + super().__init__() + if kv_input_dim != dim: + self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias) + else: + self.kv_mapper = None + + self.norm1 = RMSNorm(dim, 1e-06) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim, + out_bias=attention_out_bias, + ) + + self.norm2 = RMSNorm(dim, 1e-06) + + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + out_bias=attention_out_bias, + ) + + def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs): + cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} + + if self.kv_mapper is not None: + encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states)) + + norm_hidden_states = self.norm1(hidden_states) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + norm_hidden_states = self.norm2(hidden_states) + + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + return hidden_states + + +class FeedForward(nn.Module): + r""" + A feed-forward layer. + + Parameters: + dim (`int`): The number of channels in the input. + dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`. + mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + final_dropout (`bool` *optional*, defaults to False): Apply a final dropout. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__( + self, + dim: int, + dim_out: Optional[int] = None, + mult: int = 4, + dropout: float = 0.0, + activation_fn: str = "geglu", + final_dropout: bool = False, + inner_dim=None, + bias: bool = True, + ): + super().__init__() + if inner_dim is None: + inner_dim = int(dim * mult) + dim_out = dim_out if dim_out is not None else dim + linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear + + if activation_fn == "gelu": + act_fn = GELU(dim, inner_dim, bias=bias) + if activation_fn == "gelu-approximate": + act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias) + elif activation_fn == "geglu": + act_fn = GEGLU(dim, inner_dim, bias=bias) + elif activation_fn == "geglu-approximate": + act_fn = ApproximateGELU(dim, inner_dim, bias=bias) + + self.net = nn.ModuleList([]) + # project in + self.net.append(act_fn) + # project dropout + self.net.append(nn.Dropout(dropout)) + # project out + self.net.append(linear_cls(inner_dim, dim_out, bias=bias)) + # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout + if final_dropout: + self.net.append(nn.Dropout(dropout)) + + def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor: + compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear) + for module in self.net: + if isinstance(module, compatible_cls): + hidden_states = module(hidden_states, scale) + else: + hidden_states = module(hidden_states) + return hidden_states diff --git a/extensions-builtin/forge_space_idm_vton/src/transformerhacked_garmnet.py b/extensions-builtin/forge_space_idm_vton/src/transformerhacked_garmnet.py new file mode 100644 index 00000000..488ceb24 --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/src/transformerhacked_garmnet.py @@ -0,0 +1,460 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, Optional + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.models.embeddings import ImagePositionalEmbeddings +from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version +from src.attentionhacked_garmnet import BasicTransformerBlock +from diffusers.models.embeddings import PatchEmbed, PixArtAlphaTextProjection +from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear +from diffusers.models.modeling_utils import ModelMixin +from diffusers.models.normalization import AdaLayerNormSingle + + +@dataclass +class Transformer2DModelOutput(BaseOutput): + """ + The output of [`Transformer2DModel`]. + + Args: + sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete): + The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability + distributions for the unnoised latent pixels. + """ + + sample: torch.FloatTensor + + +class Transformer2DModel(ModelMixin, ConfigMixin): + """ + A 2D Transformer model for image-like data. + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + The number of channels in the input and output (specify if the input is **continuous**). + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use. + sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**). + This is fixed during training since it is used to learn a number of position embeddings. + num_vector_embeds (`int`, *optional*): + The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**). + Includes the class for the masked latent pixel. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): + The number of diffusion steps used during training. Pass if at least one of the norm_layers is + `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are + added to the hidden states. + + During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`. + attention_bias (`bool`, *optional*): + Configure if the `TransformerBlocks` attention should contain a bias parameter. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + out_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + patch_size: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_type: str = "layer_norm", + norm_elementwise_affine: bool = True, + norm_eps: float = 1e-5, + attention_type: str = "default", + caption_channels: int = None, + ): + super().__init__() + self.use_linear_projection = use_linear_projection + self.num_attention_heads = num_attention_heads + self.attention_head_dim = attention_head_dim + inner_dim = num_attention_heads * attention_head_dim + + conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv + linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear + + # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)` + # Define whether input is continuous or discrete depending on configuration + self.is_input_continuous = (in_channels is not None) and (patch_size is None) + self.is_input_vectorized = num_vector_embeds is not None + self.is_input_patches = in_channels is not None and patch_size is not None + + if norm_type == "layer_norm" and num_embeds_ada_norm is not None: + deprecation_message = ( + f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or" + " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config." + " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect" + " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it" + " would be very nice if you could open a Pull request for the `transformer/config.json` file" + ) + deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False) + norm_type = "ada_norm" + + if self.is_input_continuous and self.is_input_vectorized: + raise ValueError( + f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make" + " sure that either `in_channels` or `num_vector_embeds` is None." + ) + elif self.is_input_vectorized and self.is_input_patches: + raise ValueError( + f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make" + " sure that either `num_vector_embeds` or `num_patches` is None." + ) + elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches: + raise ValueError( + f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:" + f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None." + ) + + # 2. Define input layers + if self.is_input_continuous: + self.in_channels = in_channels + + self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True) + if use_linear_projection: + self.proj_in = linear_cls(in_channels, inner_dim) + else: + self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) + elif self.is_input_vectorized: + assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size" + assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed" + + self.height = sample_size + self.width = sample_size + self.num_vector_embeds = num_vector_embeds + self.num_latent_pixels = self.height * self.width + + self.latent_image_embedding = ImagePositionalEmbeddings( + num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width + ) + elif self.is_input_patches: + assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size" + + self.height = sample_size + self.width = sample_size + + self.patch_size = patch_size + interpolation_scale = self.config.sample_size // 64 # => 64 (= 512 pixart) has interpolation scale 1 + interpolation_scale = max(interpolation_scale, 1) + self.pos_embed = PatchEmbed( + height=sample_size, + width=sample_size, + patch_size=patch_size, + in_channels=in_channels, + embed_dim=inner_dim, + interpolation_scale=interpolation_scale, + ) + + # 3. Define transformers blocks + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + cross_attention_dim=cross_attention_dim, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + attention_bias=attention_bias, + only_cross_attention=only_cross_attention, + double_self_attention=double_self_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + norm_elementwise_affine=norm_elementwise_affine, + norm_eps=norm_eps, + attention_type=attention_type, + ) + for d in range(num_layers) + ] + ) + + # 4. Define output layers + self.out_channels = in_channels if out_channels is None else out_channels + if self.is_input_continuous: + # TODO: should use out_channels for continuous projections + if use_linear_projection: + self.proj_out = linear_cls(inner_dim, in_channels) + else: + self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) + elif self.is_input_vectorized: + self.norm_out = nn.LayerNorm(inner_dim) + self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1) + elif self.is_input_patches and norm_type != "ada_norm_single": + self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) + self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim) + self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) + elif self.is_input_patches and norm_type == "ada_norm_single": + self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) + self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5) + self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) + + # 5. PixArt-Alpha blocks. + self.adaln_single = None + self.use_additional_conditions = False + if norm_type == "ada_norm_single": + self.use_additional_conditions = self.config.sample_size == 128 + # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use + # additional conditions until we find better name + self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions) + + self.caption_projection = None + if caption_channels is not None: + self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim) + + self.gradient_checkpointing = False + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + timestep: Optional[torch.LongTensor] = None, + added_cond_kwargs: Dict[str, torch.Tensor] = None, + class_labels: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + ): + """ + The [`Transformer2DModel`] forward method. + + Args: + hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous): + Input `hidden_states`. + encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.LongTensor`, *optional*): + Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`. + class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*): + Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in + `AdaLayerZeroNorm`. + cross_attention_kwargs ( `Dict[str, Any]`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + attention_mask ( `torch.Tensor`, *optional*): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + encoder_attention_mask ( `torch.Tensor`, *optional*): + Cross-attention mask applied to `encoder_hidden_states`. Two formats supported: + + * Mask `(batch, sequence_length)` True = keep, False = discard. + * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard. + + If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format + above. This bias will be added to the cross-attention scores. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + + Returns: + If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a + `tuple` where the first element is the sample tensor. + """ + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension. + # we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward. + # we can tell by counting dims; if ndim == 2: it's a mask rather than a bias. + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None and attention_mask.ndim == 2: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2: + encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # Retrieve lora scale. + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + # 1. Input + if self.is_input_continuous: + batch, _, height, width = hidden_states.shape + residual = hidden_states + + hidden_states = self.norm(hidden_states) + if not self.use_linear_projection: + hidden_states = ( + self.proj_in(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_in(hidden_states) + ) + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + else: + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + hidden_states = ( + self.proj_in(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_in(hidden_states) + ) + + elif self.is_input_vectorized: + hidden_states = self.latent_image_embedding(hidden_states) + elif self.is_input_patches: + height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size + hidden_states = self.pos_embed(hidden_states) + + if self.adaln_single is not None: + if self.use_additional_conditions and added_cond_kwargs is None: + raise ValueError( + "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`." + ) + batch_size = hidden_states.shape[0] + timestep, embedded_timestep = self.adaln_single( + timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype + ) + + # 2. Blocks + if self.caption_projection is not None: + batch_size = hidden_states.shape[0] + encoder_hidden_states = self.caption_projection(encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1]) + + garment_features = [] + for block in self.transformer_blocks: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states,out_garment_feat = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + timestep, + cross_attention_kwargs, + class_labels, + **ckpt_kwargs, + ) + else: + hidden_states,out_garment_feat = block( + hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=class_labels, + ) + garment_features += out_garment_feat + # 3. Output + if self.is_input_continuous: + if not self.use_linear_projection: + hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + hidden_states = ( + self.proj_out(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_out(hidden_states) + ) + else: + hidden_states = ( + self.proj_out(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_out(hidden_states) + ) + hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + + output = hidden_states + residual + elif self.is_input_vectorized: + hidden_states = self.norm_out(hidden_states) + logits = self.out(hidden_states) + # (batch, self.num_vector_embeds - 1, self.num_latent_pixels) + logits = logits.permute(0, 2, 1) + + # log(p(x_0)) + output = F.log_softmax(logits.double(), dim=1).float() + + if self.is_input_patches: + if self.config.norm_type != "ada_norm_single": + conditioning = self.transformer_blocks[0].norm1.emb( + timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None] + hidden_states = self.proj_out_2(hidden_states) + elif self.config.norm_type == "ada_norm_single": + shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) + # Modulation + hidden_states = hidden_states * (1 + scale) + shift + hidden_states = self.proj_out(hidden_states) + hidden_states = hidden_states.squeeze(1) + + # unpatchify + if self.adaln_single is None: + height = width = int(hidden_states.shape[1] ** 0.5) + hidden_states = hidden_states.reshape( + shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels) + ) + hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) + output = hidden_states.reshape( + shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size) + ) + + if not return_dict: + return (output,) ,garment_features + + return Transformer2DModelOutput(sample=output),garment_features diff --git a/extensions-builtin/forge_space_idm_vton/src/transformerhacked_tryon.py b/extensions-builtin/forge_space_idm_vton/src/transformerhacked_tryon.py new file mode 100644 index 00000000..9a4ac5da --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/src/transformerhacked_tryon.py @@ -0,0 +1,467 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, Optional + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.models.embeddings import ImagePositionalEmbeddings +from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version +from src.attentionhacked_tryon import BasicTransformerBlock +from diffusers.models.embeddings import PatchEmbed, PixArtAlphaTextProjection +from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear +from diffusers.models.modeling_utils import ModelMixin +from diffusers.models.normalization import AdaLayerNormSingle + + +@dataclass +class Transformer2DModelOutput(BaseOutput): + """ + The output of [`Transformer2DModel`]. + + Args: + sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete): + The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability + distributions for the unnoised latent pixels. + """ + + sample: torch.FloatTensor + + +class Transformer2DModel(ModelMixin, ConfigMixin): + """ + A 2D Transformer model for image-like data. + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + The number of channels in the input and output (specify if the input is **continuous**). + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use. + sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**). + This is fixed during training since it is used to learn a number of position embeddings. + num_vector_embeds (`int`, *optional*): + The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**). + Includes the class for the masked latent pixel. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): + The number of diffusion steps used during training. Pass if at least one of the norm_layers is + `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are + added to the hidden states. + + During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`. + attention_bias (`bool`, *optional*): + Configure if the `TransformerBlocks` attention should contain a bias parameter. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + out_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + patch_size: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_type: str = "layer_norm", + norm_elementwise_affine: bool = True, + norm_eps: float = 1e-5, + attention_type: str = "default", + caption_channels: int = None, + ): + super().__init__() + self.use_linear_projection = use_linear_projection + self.num_attention_heads = num_attention_heads + self.attention_head_dim = attention_head_dim + inner_dim = num_attention_heads * attention_head_dim + + conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv + linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear + + # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)` + # Define whether input is continuous or discrete depending on configuration + self.is_input_continuous = (in_channels is not None) and (patch_size is None) + self.is_input_vectorized = num_vector_embeds is not None + self.is_input_patches = in_channels is not None and patch_size is not None + + if norm_type == "layer_norm" and num_embeds_ada_norm is not None: + deprecation_message = ( + f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or" + " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config." + " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect" + " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it" + " would be very nice if you could open a Pull request for the `transformer/config.json` file" + ) + deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False) + norm_type = "ada_norm" + + if self.is_input_continuous and self.is_input_vectorized: + raise ValueError( + f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make" + " sure that either `in_channels` or `num_vector_embeds` is None." + ) + elif self.is_input_vectorized and self.is_input_patches: + raise ValueError( + f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make" + " sure that either `num_vector_embeds` or `num_patches` is None." + ) + elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches: + raise ValueError( + f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:" + f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None." + ) + + # 2. Define input layers + if self.is_input_continuous: + self.in_channels = in_channels + + self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True) + if use_linear_projection: + self.proj_in = linear_cls(in_channels, inner_dim) + else: + self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) + elif self.is_input_vectorized: + assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size" + assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed" + + self.height = sample_size + self.width = sample_size + self.num_vector_embeds = num_vector_embeds + self.num_latent_pixels = self.height * self.width + + self.latent_image_embedding = ImagePositionalEmbeddings( + num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width + ) + elif self.is_input_patches: + assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size" + + self.height = sample_size + self.width = sample_size + + self.patch_size = patch_size + interpolation_scale = self.config.sample_size // 64 # => 64 (= 512 pixart) has interpolation scale 1 + interpolation_scale = max(interpolation_scale, 1) + self.pos_embed = PatchEmbed( + height=sample_size, + width=sample_size, + patch_size=patch_size, + in_channels=in_channels, + embed_dim=inner_dim, + interpolation_scale=interpolation_scale, + ) + + # 3. Define transformers blocks + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + cross_attention_dim=cross_attention_dim, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + attention_bias=attention_bias, + only_cross_attention=only_cross_attention, + double_self_attention=double_self_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + norm_elementwise_affine=norm_elementwise_affine, + norm_eps=norm_eps, + attention_type=attention_type, + ) + for d in range(num_layers) + ] + ) + + # 4. Define output layers + self.out_channels = in_channels if out_channels is None else out_channels + if self.is_input_continuous: + # TODO: should use out_channels for continuous projections + if use_linear_projection: + self.proj_out = linear_cls(inner_dim, in_channels) + else: + self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) + elif self.is_input_vectorized: + self.norm_out = nn.LayerNorm(inner_dim) + self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1) + elif self.is_input_patches and norm_type != "ada_norm_single": + self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) + self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim) + self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) + elif self.is_input_patches and norm_type == "ada_norm_single": + self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) + self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5) + self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) + + # 5. PixArt-Alpha blocks. + self.adaln_single = None + self.use_additional_conditions = False + if norm_type == "ada_norm_single": + self.use_additional_conditions = self.config.sample_size == 128 + # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use + # additional conditions until we find better name + self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions) + + self.caption_projection = None + if caption_channels is not None: + self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim) + + self.gradient_checkpointing = False + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + timestep: Optional[torch.LongTensor] = None, + added_cond_kwargs: Dict[str, torch.Tensor] = None, + class_labels: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + garment_features=None, + curr_garment_feat_idx=0, + return_dict: bool = True, + ): + """ + The [`Transformer2DModel`] forward method. + + Args: + hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous): + Input `hidden_states`. + encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.LongTensor`, *optional*): + Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`. + class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*): + Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in + `AdaLayerZeroNorm`. + cross_attention_kwargs ( `Dict[str, Any]`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + attention_mask ( `torch.Tensor`, *optional*): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + encoder_attention_mask ( `torch.Tensor`, *optional*): + Cross-attention mask applied to `encoder_hidden_states`. Two formats supported: + + * Mask `(batch, sequence_length)` True = keep, False = discard. + * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard. + + If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format + above. This bias will be added to the cross-attention scores. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + + Returns: + If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a + `tuple` where the first element is the sample tensor. + """ + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension. + # we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward. + # we can tell by counting dims; if ndim == 2: it's a mask rather than a bias. + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None and attention_mask.ndim == 2: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2: + encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # Retrieve lora scale. + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + # 1. Input + if self.is_input_continuous: + batch, _, height, width = hidden_states.shape + residual = hidden_states + + hidden_states = self.norm(hidden_states) + if not self.use_linear_projection: + hidden_states = ( + self.proj_in(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_in(hidden_states) + ) + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + else: + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + hidden_states = ( + self.proj_in(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_in(hidden_states) + ) + + elif self.is_input_vectorized: + hidden_states = self.latent_image_embedding(hidden_states) + elif self.is_input_patches: + height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size + hidden_states = self.pos_embed(hidden_states) + + if self.adaln_single is not None: + if self.use_additional_conditions and added_cond_kwargs is None: + raise ValueError( + "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`." + ) + batch_size = hidden_states.shape[0] + timestep, embedded_timestep = self.adaln_single( + timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype + ) + + # 2. Blocks + if self.caption_projection is not None: + batch_size = hidden_states.shape[0] + encoder_hidden_states = self.caption_projection(encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1]) + + + for block in self.transformer_blocks: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states,curr_garment_feat_idx = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + timestep, + cross_attention_kwargs, + class_labels, + garment_features, + curr_garment_feat_idx, + **ckpt_kwargs, + ) + else: + hidden_states,curr_garment_feat_idx = block( + hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=class_labels, + garment_features=garment_features, + curr_garment_feat_idx=curr_garment_feat_idx, + ) + + + # 3. Output + if self.is_input_continuous: + if not self.use_linear_projection: + hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + hidden_states = ( + self.proj_out(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_out(hidden_states) + ) + else: + hidden_states = ( + self.proj_out(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_out(hidden_states) + ) + hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + + output = hidden_states + residual + elif self.is_input_vectorized: + hidden_states = self.norm_out(hidden_states) + logits = self.out(hidden_states) + # (batch, self.num_vector_embeds - 1, self.num_latent_pixels) + logits = logits.permute(0, 2, 1) + + # log(p(x_0)) + output = F.log_softmax(logits.double(), dim=1).float() + + if self.is_input_patches: + if self.config.norm_type != "ada_norm_single": + conditioning = self.transformer_blocks[0].norm1.emb( + timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None] + hidden_states = self.proj_out_2(hidden_states) + elif self.config.norm_type == "ada_norm_single": + shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) + # Modulation + hidden_states = hidden_states * (1 + scale) + shift + hidden_states = self.proj_out(hidden_states) + hidden_states = hidden_states.squeeze(1) + + # unpatchify + if self.adaln_single is None: + height = width = int(hidden_states.shape[1] ** 0.5) + hidden_states = hidden_states.reshape( + shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels) + ) + hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) + output = hidden_states.reshape( + shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size) + ) + + if not return_dict: + return (output,),curr_garment_feat_idx + + return Transformer2DModelOutput(sample=output),curr_garment_feat_idx diff --git a/extensions-builtin/forge_space_idm_vton/src/tryon_pipeline.py b/extensions-builtin/forge_space_idm_vton/src/tryon_pipeline.py new file mode 100644 index 00000000..78f22375 --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/src/tryon_pipeline.py @@ -0,0 +1,1893 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from diffusers.image_processor import PipelineImageInput, VaeImageProcessor +from diffusers.loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from diffusers.models.attention_processor import ( + AttnProcessor2_0, + FusedAttnProcessor2_0, + LoRAAttnProcessor2_0, + LoRAXFormersAttnProcessor, + XFormersAttnProcessor, +) +from diffusers.models.lora import adjust_lora_scale_text_encoder +from diffusers.schedulers import KarrasDiffusionSchedulers +from diffusers.utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from diffusers.utils.torch_utils import randn_tensor +from diffusers.pipelines.pipeline_utils import DiffusionPipeline + + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLInpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = StableDiffusionXLInpaintPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", + ... torch_dtype=torch.float16, + ... variant="fp16", + ... use_safetensors=True, + ... ) + >>> pipe.to("cuda") + + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + + >>> init_image = load_image(img_url).convert("RGB") + >>> mask_image = load_image(mask_url).convert("RGB") + + >>> prompt = "A majestic tiger sitting on a bench" + >>> image = pipe( + ... prompt=prompt, image=init_image, mask_image=mask_image, num_inference_steps=50, strength=0.80 + ... ).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +def mask_pil_to_torch(mask, height, width): + # preprocess mask + if isinstance(mask, (PIL.Image.Image, np.ndarray)): + mask = [mask] + + if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image): + mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask] + mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0) + mask = mask.astype(np.float32) / 255.0 + elif isinstance(mask, list) and isinstance(mask[0], np.ndarray): + mask = np.concatenate([m[None, None, :] for m in mask], axis=0) + + mask = torch.from_numpy(mask) + return mask + + +def prepare_mask_and_masked_image(image, mask, height, width, return_image: bool = False): + """ + Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be + converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the + ``image`` and ``1`` for the ``mask``. + + The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be + binarized (``mask > 0.5``) and cast to ``torch.float32`` too. + + Args: + image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint. + It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width`` + ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``. + mask (_type_): The mask to apply to the image, i.e. regions to inpaint. + It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width`` + ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``. + + + Raises: + ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask + should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions. + TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not + (ot the other way around). + + Returns: + tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4 + dimensions: ``batch x channels x height x width``. + """ + + # checkpoint. TOD(Yiyi) - need to clean this up later + deprecation_message = "The prepare_mask_and_masked_image method is deprecated and will be removed in a future version. Please use VaeImageProcessor.preprocess instead" + deprecate( + "prepare_mask_and_masked_image", + "0.30.0", + deprecation_message, + ) + if image is None: + raise ValueError("`image` input cannot be undefined.") + + if mask is None: + raise ValueError("`mask_image` input cannot be undefined.") + + if isinstance(image, torch.Tensor): + if not isinstance(mask, torch.Tensor): + mask = mask_pil_to_torch(mask, height, width) + + if image.ndim == 3: + image = image.unsqueeze(0) + + # Batch and add channel dim for single mask + if mask.ndim == 2: + mask = mask.unsqueeze(0).unsqueeze(0) + + # Batch single mask or add channel dim + if mask.ndim == 3: + # Single batched mask, no channel dim or single mask not batched but channel dim + if mask.shape[0] == 1: + mask = mask.unsqueeze(0) + + # Batched masks no channel dim + else: + mask = mask.unsqueeze(1) + + assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions" + # assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions" + assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size" + + # Check image is in [-1, 1] + # if image.min() < -1 or image.max() > 1: + # raise ValueError("Image should be in [-1, 1] range") + + # Check mask is in [0, 1] + if mask.min() < 0 or mask.max() > 1: + raise ValueError("Mask should be in [0, 1] range") + + # Binarize mask + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + + # Image as float32 + image = image.to(dtype=torch.float32) + elif isinstance(mask, torch.Tensor): + raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not") + else: + # preprocess image + if isinstance(image, (PIL.Image.Image, np.ndarray)): + image = [image] + if isinstance(image, list) and isinstance(image[0], PIL.Image.Image): + # resize all images w.r.t passed height an width + image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image] + image = [np.array(i.convert("RGB"))[None, :] for i in image] + image = np.concatenate(image, axis=0) + elif isinstance(image, list) and isinstance(image[0], np.ndarray): + image = np.concatenate([i[None, :] for i in image], axis=0) + + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + + mask = mask_pil_to_torch(mask, height, width) + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + + if image.shape[1] == 4: + # images are in latent space and thus can't + # be masked set masked_image to None + # we assume that the checkpoint is not an inpainting + # checkpoint. TOD(Yiyi) - need to clean this up later + masked_image = None + else: + masked_image = image * (mask < 0.5) + + # n.b. ensure backwards compatibility as old function does not return image + if return_image: + return mask, masked_image, image + + return mask, masked_image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + **kwargs, +): + """ + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps` + must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLInpaintPipeline( + DiffusionPipeline, + TextualInversionLoaderMixin, + StableDiffusionXLLoraLoaderMixin, + FromSingleFileMixin, + IPAdapterMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`): + Whether the `unet` requires a aesthetic_score condition to be passed during inference. Also see the config + of `stabilityai/stable-diffusion-xl-refiner-1-0`. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "add_neg_time_ids", + "mask", + "masked_image_latents", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + requires_aesthetics_score: bool = False, + force_zeros_for_empty_prompt: bool = True, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + scheduler=scheduler, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.register_to_config(requires_aesthetics_score=requires_aesthetics_score) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + + + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + # print(image.shape) + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds(self, ip_adapter_image, device, num_images_per_prompt): + # if not isinstance(ip_adapter_image, list): + # ip_adapter_image = [ip_adapter_image] + + # if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + # raise ValueError( + # f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + # ) + output_hidden_state = not isinstance(self.unet.encoder_hid_proj, ImageProjection) + # print(output_hidden_state) + image_embeds, negative_image_embeds = self.encode_image( + ip_adapter_image, device, 1, output_hidden_state + ) + # print(single_image_embeds.shape) + # single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0) + # single_negative_image_embeds = torch.stack([single_negative_image_embeds] * num_images_per_prompt, dim=0) + # print(single_image_embeds.shape) + if self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds]) + image_embeds = image_embeds.to(device) + + + return image_embeds + + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: procecss multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + pooled_prompt_embeds = prompt_embeds[0] + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (ฮท) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to ฮท in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.") + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=True, + add_noise=True, + return_noise=False, + return_image_latents=False, + ): + shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if (image is None or timestep is None) and not is_strength_max: + raise ValueError( + "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." + "However, either the image or the noise timestep has not been provided." + ) + + if image.shape[1] == 4: + image_latents = image.to(device=device, dtype=dtype) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + elif return_image_latents or (latents is None and not is_strength_max): + image = image.to(device=device, dtype=dtype) + image_latents = self._encode_vae_image(image=image, generator=generator) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + + if latents is None and add_noise: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep) + # if pure noise then scale the initial latents by the Scheduler's init sigma + latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents + elif add_noise: + noise = latents.to(device) + latents = noise * self.scheduler.init_noise_sigma + else: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = image_latents.to(device) + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + dtype = image.dtype + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + image_latents = image_latents.to(dtype) + image_latents = self.vae.config.scaling_factor * image_latents + + return image_latents + + def prepare_mask_latents( + self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate( + mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) + ) + mask = mask.to(device=device, dtype=dtype) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + if masked_image is not None and masked_image.shape[1] == 4: + masked_image_latents = masked_image + else: + masked_image_latents = None + + if masked_image is not None: + if masked_image_latents is None: + masked_image = masked_image.to(device=device, dtype=dtype) + masked_image_latents = self._encode_vae_image(masked_image, generator=generator) + + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat( + batch_size // masked_image_latents.shape[0], 1, 1, 1 + ) + + masked_image_latents = ( + torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + + return mask, masked_image_latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + else: + t_start = 0 + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + if denoising_start is not None: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + + num_inference_steps = (timesteps < discrete_timestep_cutoff).sum().item() + if self.scheduler.order == 2 and num_inference_steps % 2 == 0: + # if the scheduler is a 2nd order scheduler we might have to do +1 + # because `num_inference_steps` might be even given that every timestep + # (except the highest one) is duplicated. If `num_inference_steps` is even it would + # mean that we cut the timesteps in the middle of the denoising step + # (between 1st and 2nd devirative) which leads to incorrect results. By adding 1 + # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler + num_inference_steps = num_inference_steps + 1 + + # because t_n+1 >= t_n, we slice the timesteps starting from the end + timesteps = timesteps[-num_inference_steps:] + return timesteps, num_inference_steps + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids + def _get_add_time_ids( + self, + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype, + text_encoder_projection_dim=None, + ): + if self.config.requires_aesthetics_score: + add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,)) + add_neg_time_ids = list( + negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,) + ) + else: + add_time_ids = list(original_size + crops_coords_top_left + target_size) + add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if ( + expected_add_embed_dim > passed_add_embed_dim + and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model." + ) + elif ( + expected_add_embed_dim < passed_add_embed_dim + and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model." + ) + elif expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype) + + return add_time_ids, add_neg_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + LoRAXFormersAttnProcessor, + LoRAAttnProcessor2_0, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu + def enable_freeu(self, s1: float, s2: float, b1: float, b2: float): + r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497. + + The suffixes after the scaling factors represent the stages where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values + that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + if not hasattr(self, "unet"): + raise ValueError("The pipeline must have `unet` for using FreeU.") + self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu + def disable_freeu(self): + """Disables the FreeU mechanism if enabled.""" + self.unet.disable_freeu() + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections + def fuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is ๐Ÿงช experimental. + + + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + """ + self.fusing_unet = False + self.fusing_vae = False + + if unet: + self.fusing_unet = True + self.unet.fuse_qkv_projections() + self.unet.set_attn_processor(FusedAttnProcessor2_0()) + + if vae: + if not isinstance(self.vae, AutoencoderKL): + raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.") + + self.fusing_vae = True + self.vae.fuse_qkv_projections() + self.vae.set_attn_processor(FusedAttnProcessor2_0()) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections + def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """Disable QKV projection fusion if enabled. + + + + This API is ๐Ÿงช experimental. + + + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + + """ + if unet: + if not self.fusing_unet: + logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.") + else: + self.unet.unfuse_qkv_projections() + self.fusing_unet = False + + if vae: + if not self.fusing_vae: + logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.") + else: + self.vae.unfuse_qkv_projections() + self.fusing_vae = False + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32): + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + timesteps (`torch.Tensor`): + generate embedding vectors at these timesteps + embedding_dim (`int`, *optional*, defaults to 512): + dimension of the embeddings to generate + dtype: + data type of the generated embeddings + + Returns: + `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)` + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def denoising_start(self): + return self._denoising_start + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: torch.FloatTensor = None, + height: Optional[int] = None, + width: Optional[int] = None, + padding_mask_crop: Optional[int] = None, + strength: float = 0.9999, + num_inference_steps: int = 50, + timesteps: List[int] = None, + denoising_start: Optional[float] = None, + denoising_end: Optional[float] = None, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + output_type: Optional[str] = "pil", + cloth =None, + pose_img = None, + text_embeds_cloth=None, + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + original_size: Tuple[int, int] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Tuple[int, int] = None, + negative_original_size: Optional[Tuple[int, int]] = None, + negative_crops_coords_top_left: Tuple[int, int] = (0, 0), + negative_target_size: Optional[Tuple[int, int]] = None, + aesthetic_score: float = 6.0, + negative_aesthetic_score: float = 2.5, + clip_skip: Optional[int] = None, + pooled_prompt_embeds_c=None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will + be masked out with `mask_image` and repainted according to `prompt`. + mask_image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted + to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L) + instead of 3, so the expected shape would be `(B, H, W, 1)`. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to image and mask_image. If + `padding_mask_crop` is not `None`, it will first find a rectangular region with the same aspect ration of the image and + contains all masked area, and then expand that area based on `padding_mask_crop`. The image and mask_image will then be cropped based on + the expanded area before resizing to the original image size for inpainting. This is useful when the masked area is small while the image is large + and contain information inreleant for inpainging, such as background. + strength (`float`, *optional*, defaults to 0.9999): + Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be + between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the + `strength`. The number of denoising steps depends on the amount of noise initially added. When + `strength` is 1, added noise will be maximum and the denoising process will run for the full number of + iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores the masked + portion of the reference `image`. Note that in the case of `denoising_start` being declared as an + integer, the value of `strength` will be ignored. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + denoising_start (`float`, *optional*): + When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be + bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and + it is assumed that the passed `image` is a partly denoised image. Note that when this is specified, + strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline + is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output). + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be + denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the + final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline + forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output). + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (ฮท) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + aesthetic_score (`float`, *optional*, defaults to 6.0): + Used to simulate an aesthetic score of the generated image by influencing the positive text condition. + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_aesthetic_score (`float`, *optional*, defaults to 2.5): + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to + simulate an aesthetic score of the generated image by influencing the negative text condition. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple. `tuple. When returning a tuple, the first element is a list with the generated images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs + self.check_inputs( + prompt, + prompt_2, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + padding_mask_crop, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._denoising_start = denoising_start + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. set timesteps + def denoising_value_valid(dnv): + return isinstance(self.denoising_end, float) and 0 < dnv < 1 + + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, + strength, + device, + denoising_start=self.denoising_start if denoising_value_valid else None, + ) + # check that number of inference steps is not < 1 - as this doesn't make sense + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + # at which timestep to set the initial noise (n.b. 50% if strength is 0.5) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + + # 5. Preprocess mask and image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + mask = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + if masked_image_latents is not None: + masked_image = masked_image_latents + elif init_image.shape[1] == 4: + # if images are in latent space, we can't mask it + masked_image = None + else: + masked_image = init_image * (mask < 0.5) + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_unet = self.unet.config.in_channels + return_image_latents = num_channels_unet == 4 + + add_noise = True if self.denoising_start is None else False + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image=init_image, + timestep=latent_timestep, + is_strength_max=is_strength_max, + add_noise=add_noise, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 7. Prepare mask latent variables + mask, masked_image_latents = self.prepare_mask_latents( + mask, + masked_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + pose_img = pose_img.to(device=device, dtype=prompt_embeds.dtype) + + pose_img = self.vae.encode(pose_img).latent_dist.sample() + pose_img = pose_img * self.vae.config.scaling_factor + + # pose_img = self._encode_vae_image(pose_img, generator=generator) + + pose_img = ( + torch.cat([pose_img] * 2) if self.do_classifier_free_guidance else pose_img + ) + cloth = self._encode_vae_image(cloth, generator=generator) + + # # 8. Check that sizes of mask, masked image and latents match + # if num_channels_unet == 9: + # # default case for runwayml/stable-diffusion-inpainting + # num_channels_mask = mask.shape[1] + # num_channels_masked_image = masked_image_latents.shape[1] + # if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels: + # raise ValueError( + # f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + # f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + # f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + # f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of" + # " `pipeline.unet` or your `mask_image` or `image` input." + # ) + # elif num_channels_unet != 4: + # raise ValueError( + # f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}." + # ) + # 8.1 Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + height, width = latents.shape[-2:] + height = height * self.vae_scale_factor + width = width * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 10. Prepare added time ids & embeddings + if negative_original_size is None: + negative_original_size = original_size + if negative_target_size is None: + negative_target_size = target_size + + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids, add_neg_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1) + add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device) + + if ip_adapter_image is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, device, batch_size * num_images_per_prompt + ) + + #project outside for loop + image_embeds = self.unet.encoder_hid_proj(image_embeds).to(prompt_embeds.dtype) + + + # 11. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + if ( + self.denoising_end is not None + and self.denoising_start is not None + and denoising_value_valid(self.denoising_end) + and denoising_value_valid(self.denoising_start) + and self.denoising_start >= self.denoising_end + ): + raise ValueError( + f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: " + + f" {self.denoising_end} when using type float." + ) + elif self.denoising_end is not None and denoising_value_valid(self.denoising_end): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 11.1 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # concat latents, mask, masked_image_latents in the channel dimension + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + + # bsz = mask.shape[0] + if num_channels_unet == 13: + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents,pose_img], dim=1) + + # if num_channels_unet == 9: + # latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image is not None: + added_cond_kwargs["image_embeds"] = image_embeds + # down,reference_features = self.UNet_Encoder(cloth,t, text_embeds_cloth,added_cond_kwargs= {"text_embeds": pooled_prompt_embeds_c, "time_ids": add_time_ids},return_dict=False) + down,reference_features = self.unet_encoder(cloth,t, text_embeds_cloth,return_dict=False) + # print(type(reference_features)) + # print(reference_features) + reference_features = list(reference_features) + # print(len(reference_features)) + # for elem in reference_features: + # print(elem.shape) + # exit(1) + if self.do_classifier_free_guidance: + reference_features = [torch.cat([torch.zeros_like(d), d]) for d in reference_features] + + + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + garment_features=reference_features, + )[0] + # noise_pred = self.unet(latent_model_input, t, + # prompt_embeds,timestep_cond=timestep_cond,cross_attention_kwargs=self.cross_attention_kwargs,added_cond_kwargs=added_cond_kwargs,down_block_additional_attn=down ).sample + + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if num_channels_unet == 4: + init_latents_proper = image_latents + if self.do_classifier_free_guidance: + init_mask, _ = mask.chunk(2) + else: + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.add_noise( + init_latents_proper, noise, torch.tensor([noise_timestep]) + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + add_neg_time_ids = callback_outputs.pop("add_neg_time_ids", add_neg_time_ids) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + # else: + # return StableDiffusionXLPipelineOutput(images=latents) + + + image = self.image_processor.postprocess(image, output_type=output_type) + + if padding_mask_crop is not None: + image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image] + + # Offload all models + self.maybe_free_model_hooks() + + # if not return_dict: + return (image,) + + # return StableDiffusionXLPipelineOutput(images=image) \ No newline at end of file diff --git a/extensions-builtin/forge_space_idm_vton/src/unet_block_hacked_garmnet.py b/extensions-builtin/forge_space_idm_vton/src/unet_block_hacked_garmnet.py new file mode 100644 index 00000000..a6e80c05 --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/src/unet_block_hacked_garmnet.py @@ -0,0 +1,3579 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Any, Dict, Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import is_torch_version, logging +from diffusers.utils.torch_utils import apply_freeu +from diffusers.models.activations import get_activation +from diffusers.models.attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0 +from diffusers.models.dual_transformer_2d import DualTransformer2DModel +from diffusers.models.normalization import AdaGroupNorm +from diffusers.models.resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D +from src.transformerhacked_garmnet import Transformer2DModel +from einops import rearrange + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def get_down_block( + down_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + temb_channels: int, + add_downsample: bool, + resnet_eps: float, + resnet_act_fn: str, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + downsample_padding: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + downsample_type: Optional[str] = None, + dropout: float = 0.0, +): + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warn( + f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type + if down_block_type == "DownBlock2D": + return DownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "ResnetDownsampleBlock2D": + return ResnetDownsampleBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + ) + elif down_block_type == "AttnDownBlock2D": + if add_downsample is False: + downsample_type = None + else: + downsample_type = downsample_type or "conv" # default to 'conv' + return AttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + downsample_type=downsample_type, + ) + elif down_block_type == "CrossAttnDownBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D") + return CrossAttnDownBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + elif down_block_type == "SimpleCrossAttnDownBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D") + return SimpleCrossAttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif down_block_type == "SkipDownBlock2D": + return SkipDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "AttnSkipDownBlock2D": + return AttnSkipDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "DownEncoderBlock2D": + return DownEncoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "AttnDownEncoderBlock2D": + return AttnDownEncoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "KDownBlock2D": + return KDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + ) + elif down_block_type == "KCrossAttnDownBlock2D": + return KCrossAttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + add_self_attention=True if not add_downsample else False, + ) + raise ValueError(f"{down_block_type} does not exist.") + + +def get_up_block( + up_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + add_upsample: bool, + resnet_eps: float, + resnet_act_fn: str, + resolution_idx: Optional[int] = None, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + upsample_type: Optional[str] = None, + dropout: float = 0.0, +) -> nn.Module: + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warn( + f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type + if up_block_type == "UpBlock2D": + return UpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "ResnetUpsampleBlock2D": + return ResnetUpsampleBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + ) + elif up_block_type == "CrossAttnUpBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D") + return CrossAttnUpBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + elif up_block_type == "SimpleCrossAttnUpBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D") + return SimpleCrossAttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif up_block_type == "AttnUpBlock2D": + if add_upsample is False: + upsample_type = None + else: + upsample_type = upsample_type or "conv" # default to 'conv' + + return AttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + upsample_type=upsample_type, + ) + elif up_block_type == "SkipUpBlock2D": + return SkipUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "AttnSkipUpBlock2D": + return AttnSkipUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "UpDecoderBlock2D": + return UpDecoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + temb_channels=temb_channels, + ) + elif up_block_type == "AttnUpDecoderBlock2D": + return AttnUpDecoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + temb_channels=temb_channels, + ) + elif up_block_type == "KUpBlock2D": + return KUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + ) + elif up_block_type == "KCrossAttnUpBlock2D": + return KCrossAttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + ) + + raise ValueError(f"{up_block_type} does not exist.") + + +class AutoencoderTinyBlock(nn.Module): + """ + Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU + blocks. + + Args: + in_channels (`int`): The number of input channels. + out_channels (`int`): The number of output channels. + act_fn (`str`): + ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`. + + Returns: + `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to + `out_channels`. + """ + + def __init__(self, in_channels: int, out_channels: int, act_fn: str): + super().__init__() + act_fn = get_activation(act_fn) + self.conv = nn.Sequential( + nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1), + act_fn, + nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1), + act_fn, + nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1), + ) + self.skip = ( + nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False) + if in_channels != out_channels + else nn.Identity() + ) + self.fuse = nn.ReLU() + + def forward(self, x: torch.FloatTensor) -> torch.FloatTensor: + return self.fuse(self.conv(x) + self.skip(x)) + + +class UNetMidBlock2D(nn.Module): + """ + A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks. + + Args: + in_channels (`int`): The number of input channels. + temb_channels (`int`): The number of temporal embedding channels. + dropout (`float`, *optional*, defaults to 0.0): The dropout rate. + num_layers (`int`, *optional*, defaults to 1): The number of residual blocks. + resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks. + resnet_time_scale_shift (`str`, *optional*, defaults to `default`): + The type of normalization to apply to the time embeddings. This can help to improve the performance of the + model on tasks with long-range temporal dependencies. + resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks. + resnet_groups (`int`, *optional*, defaults to 32): + The number of groups to use in the group normalization layers of the resnet blocks. + attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks. + resnet_pre_norm (`bool`, *optional*, defaults to `True`): + Whether to use pre-normalization for the resnet blocks. + add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks. + attention_head_dim (`int`, *optional*, defaults to 1): + Dimension of a single attention head. The number of attention heads is determined based on this value and + the number of input channels. + output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor. + + Returns: + `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size, + in_channels, height, width)`. + + """ + + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + attn_groups: Optional[int] = None, + resnet_pre_norm: bool = True, + add_attention: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + ): + super().__init__() + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + self.add_attention = add_attention + + if attn_groups is None: + attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}." + ) + attention_head_dim = in_channels + + for _ in range(num_layers): + if self.add_attention: + attentions.append( + Attention( + in_channels, + heads=in_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=attn_groups, + spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + else: + attentions.append(None) + + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor: + hidden_states = self.resnets[0](hidden_states, temb) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if attn is not None: + hidden_states = attn(hidden_states, temb=temb) + hidden_states = resnet(hidden_states, temb) + + return hidden_states + + +class UNetMidBlock2DCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + # support for variable transformer layers per block + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + for i in range(num_layers): + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + in_channels // num_attention_heads, + in_channels=in_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + in_channels // num_attention_heads, + in_channels=in_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale) + garment_features = [] + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + # hidden_states = attn( + hidden_states,out_garment_feat = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + ) + hidden_states=hidden_states[0] + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + else: + # hidden_states= attn( + hidden_states,out_garment_feat = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + ) + hidden_states=hidden_states[0] + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + garment_features += out_garment_feat + return hidden_states,garment_features + # return hidden_states + + +class UNetMidBlock2DSimpleCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + + self.has_cross_attention = True + + self.attention_head_dim = attention_head_dim + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + self.num_heads = in_channels // self.attention_head_dim + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ] + attentions = [] + + for _ in range(num_layers): + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=in_channels, + cross_attention_dim=in_channels, + heads=self.num_heads, + dim_head=self.attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + # attn + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + # resnet + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class AttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + downsample_type: str = "conv", + ): + super().__init__() + resnets = [] + attentions = [] + self.downsample_type = downsample_type + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if downsample_type == "conv": + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + elif downsample_type == "resnet": + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + cross_attention_kwargs.update({"scale": lora_scale}) + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn(hidden_states, **cross_attention_kwargs) + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + if self.downsample_type == "resnet": + hidden_states = downsampler(hidden_states, temb=temb, scale=lora_scale) + else: + hidden_states = downsampler(hidden_states, scale=lora_scale) + + output_states += (hidden_states,) + + return hidden_states, output_states + + +class CrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + add_downsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + additional_residuals: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + blocks = list(zip(self.resnets, self.attentions)) + garment_features = [] + for i, (resnet, attn) in enumerate(blocks): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states,out_garment_feat = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + ) + hidden_states=hidden_states[0] + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states,out_garment_feat = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + ) + hidden_states=hidden_states[0] + garment_features += out_garment_feat + # apply additional residuals to the output of the last pair of resnet and attention blocks + if i == len(blocks) - 1 and additional_residuals is not None: + hidden_states = hidden_states + additional_residuals + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale=lora_scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states,garment_features + + +class DownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale=scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class DownEncoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor: + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=None, scale=scale) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale) + + return hidden_states + + +class AttnDownEncoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb=None, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale) + + return hidden_states + + +class AttnSkipDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = np.sqrt(2.0), + add_downsample: bool = True, + ): + super().__init__() + self.attentions = nn.ModuleList([]) + self.resnets = nn.ModuleList([]) + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + self.resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(in_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + self.attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=32, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + if add_downsample: + self.resnet_down = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + down=True, + kernel="fir", + ) + self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)]) + self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1)) + else: + self.resnet_down = None + self.downsamplers = None + self.skip_conv = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + skip_sample: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]: + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + output_states += (hidden_states,) + + if self.downsamplers is not None: + hidden_states = self.resnet_down(hidden_states, temb, scale=scale) + for downsampler in self.downsamplers: + skip_sample = downsampler(skip_sample) + + hidden_states = self.skip_conv(skip_sample) + hidden_states + + output_states += (hidden_states,) + + return hidden_states, output_states, skip_sample + + +class SkipDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + output_scale_factor: float = np.sqrt(2.0), + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + self.resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(in_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + if add_downsample: + self.resnet_down = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + down=True, + kernel="fir", + ) + self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)]) + self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1)) + else: + self.resnet_down = None + self.downsamplers = None + self.skip_conv = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + skip_sample: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]: + output_states = () + + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb, scale) + output_states += (hidden_states,) + + if self.downsamplers is not None: + hidden_states = self.resnet_down(hidden_states, temb, scale) + for downsampler in self.downsamplers: + skip_sample = downsampler(skip_sample) + + hidden_states = self.skip_conv(skip_sample) + hidden_states + + output_states += (hidden_states,) + + return hidden_states, output_states, skip_sample + + +class ResnetDownsampleBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + skip_time_act: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, temb, scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class SimpleCrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + + self.has_cross_attention = True + + resnets = [] + attentions = [] + + self.attention_head_dim = attention_head_dim + self.num_heads = out_channels // self.attention_head_dim + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=out_channels, + cross_attention_dim=out_channels, + heads=self.num_heads, + dim_head=attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, temb, scale=lora_scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class KDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: int = 32, + add_downsample: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + temb_channels=temb_channels, + groups=groups, + groups_out=groups_out, + eps=resnet_eps, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + # YiYi's comments- might be able to use FirDownsample2D, look into details later + self.downsamplers = nn.ModuleList([KDownsample2D()]) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale) + + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + return hidden_states, output_states + + +class KCrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + cross_attention_dim: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_group_size: int = 32, + add_downsample: bool = True, + attention_head_dim: int = 64, + add_self_attention: bool = False, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + temb_channels=temb_channels, + groups=groups, + groups_out=groups_out, + eps=resnet_eps, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + attentions.append( + KAttentionBlock( + out_channels, + out_channels // attention_head_dim, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + temb_channels=temb_channels, + attention_bias=True, + add_self_attention=add_self_attention, + cross_attention_norm="layer_norm", + group_size=resnet_group_size, + ) + ) + + self.resnets = nn.ModuleList(resnets) + self.attentions = nn.ModuleList(attentions) + + if add_downsample: + self.downsamplers = nn.ModuleList([KDownsample2D()]) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + + if self.downsamplers is None: + output_states += (None,) + else: + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + return hidden_states, output_states + + +class AttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: int = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + upsample_type: str = "conv", + ): + super().__init__() + resnets = [] + attentions = [] + + self.upsample_type = upsample_type + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if upsample_type == "conv": + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + elif upsample_type == "resnet": + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + if self.upsample_type == "resnet": + hidden_states = upsampler(hidden_states, temb=temb, scale=scale) + else: + hidden_states = upsampler(hidden_states, scale=scale) + + return hidden_states + + +class CrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + garment_features = [] + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + # print("h.shape") + # print(h.shape) + # print("hidden_states.shape) + # print(hidden_states.shape) + # print("attn_block") + # print(attn) + + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + # print(hidden_states.shape) + # print(encoder_hidden_states.shape) + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states,out_garment_feat = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + ) + hidden_states=hidden_states[0] + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states,out_garment_feat = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + ) + hidden_states=hidden_states[0] + garment_features += out_garment_feat + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale) + + return hidden_states,garment_features + + +class UpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size, scale=scale) + + return hidden_states + # def forward( + # self, + # hidden_states: torch.FloatTensor, + # res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + # temb: Optional[torch.FloatTensor] = None, + # upsample_size: Optional[int] = None, + # scale: float = 1.0, + # zero_block=None, + # hint=None, + # ) -> torch.FloatTensor: + # is_freeu_enabled = ( + # getattr(self, "s1", None) + # and getattr(self, "s2", None) + # and getattr(self, "b1", None) + # and getattr(self, "b2", None) + # ) + + # # print(len(self.resnets)) + # # print(len(zero_block)) + # # print(len(hint)) + # # for resnet in self.resnets: + # for resnet, zero,h in zip(self.resnets,zero_block,hint): + + # # pop res hidden states + # res_hidden_states = res_hidden_states_tuple[-1] + # res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # res_hidden_states = res_hidden_states + zero(h) + # # FreeU: Only operate on the first two stages + # if is_freeu_enabled: + # hidden_states, res_hidden_states = apply_freeu( + # self.resolution_idx, + # hidden_states, + # res_hidden_states, + # s1=self.s1, + # s2=self.s2, + # b1=self.b1, + # b2=self.b2, + # ) + + # # print(hidden_states.shape) + # # # print(h.shape) + # # print(res_hidden_states.shape) + # hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + # # print(hidden_states.shape) + + # if self.training and self.gradient_checkpointing: + + # def create_custom_forward(module): + # def custom_forward(*inputs): + # return module(*inputs) + + # return custom_forward + + # if is_torch_version(">=", "1.11.0"): + # hidden_states = torch.utils.checkpoint.checkpoint( + # create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + # ) + # else: + # hidden_states = torch.utils.checkpoint.checkpoint( + # create_custom_forward(resnet), hidden_states, temb + # ) + # else: + # hidden_states = resnet(hidden_states, temb, scale=scale) + + # if self.upsamplers is not None: + # for upsampler in self.upsamplers: + # hidden_states = upsampler(hidden_states, upsample_size, scale=scale) + + # return hidden_states + + +class UpDecoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + temb_channels: Optional[int] = None, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + input_channels = in_channels if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> torch.FloatTensor: + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +class AttnUpDecoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + temb_channels: Optional[int] = None, + ): + super().__init__() + resnets = [] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + input_channels = in_channels if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None, + spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb=temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, temb=temb, **cross_attention_kwargs) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, scale=scale) + + return hidden_states + + +class AttnSkipUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = np.sqrt(2.0), + add_upsample: bool = True, + ): + super().__init__() + self.attentions = nn.ModuleList([]) + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + self.resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(resnet_in_channels + res_skip_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + self.attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=32, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels) + if add_upsample: + self.resnet_up = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + up=True, + kernel="fir", + ) + self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + self.skip_norm = torch.nn.GroupNorm( + num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True + ) + self.act = nn.SiLU() + else: + self.resnet_up = None + self.skip_conv = None + self.skip_norm = None + self.act = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + skip_sample=None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, torch.FloatTensor]: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + + cross_attention_kwargs = {"scale": scale} + hidden_states = self.attentions[0](hidden_states, **cross_attention_kwargs) + + if skip_sample is not None: + skip_sample = self.upsampler(skip_sample) + else: + skip_sample = 0 + + if self.resnet_up is not None: + skip_sample_states = self.skip_norm(hidden_states) + skip_sample_states = self.act(skip_sample_states) + skip_sample_states = self.skip_conv(skip_sample_states) + + skip_sample = skip_sample + skip_sample_states + + hidden_states = self.resnet_up(hidden_states, temb, scale=scale) + + return hidden_states, skip_sample + + +class SkipUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + output_scale_factor: float = np.sqrt(2.0), + add_upsample: bool = True, + upsample_padding: int = 1, + ): + super().__init__() + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + self.resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min((resnet_in_channels + res_skip_channels) // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels) + if add_upsample: + self.resnet_up = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + up=True, + kernel="fir", + ) + self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + self.skip_norm = torch.nn.GroupNorm( + num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True + ) + self.act = nn.SiLU() + else: + self.resnet_up = None + self.skip_conv = None + self.skip_norm = None + self.act = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + skip_sample=None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, torch.FloatTensor]: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + + if skip_sample is not None: + skip_sample = self.upsampler(skip_sample) + else: + skip_sample = 0 + + if self.resnet_up is not None: + skip_sample_states = self.skip_norm(hidden_states) + skip_sample_states = self.act(skip_sample_states) + skip_sample_states = self.skip_conv(skip_sample_states) + + skip_sample = skip_sample + skip_sample_states + + hidden_states = self.resnet_up(hidden_states, temb, scale=scale) + + return hidden_states, skip_sample + + +class ResnetUpsampleBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + skip_time_act: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, temb, scale=scale) + + return hidden_states + + +class SimpleCrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.attention_head_dim = attention_head_dim + + self.num_heads = out_channels // self.attention_head_dim + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=out_channels, + cross_attention_dim=out_channels, + heads=self.num_heads, + dim_head=self.attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + for resnet, attn in zip(self.resnets, self.attentions): + # resnet + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class KUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + resolution_idx: int, + dropout: float = 0.0, + num_layers: int = 5, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: Optional[int] = 32, + add_upsample: bool = True, + ): + super().__init__() + resnets = [] + k_in_channels = 2 * out_channels + k_out_channels = in_channels + num_layers = num_layers - 1 + + for i in range(num_layers): + in_channels = k_in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=k_out_channels if (i == num_layers - 1) else out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=groups, + groups_out=groups_out, + dropout=dropout, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([KUpsample2D()]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + res_hidden_states_tuple = res_hidden_states_tuple[-1] + if res_hidden_states_tuple is not None: + hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1) + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +class KCrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + resolution_idx: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: int = 32, + attention_head_dim: int = 1, # attention dim_head + cross_attention_dim: int = 768, + add_upsample: bool = True, + upcast_attention: bool = False, + ): + super().__init__() + resnets = [] + attentions = [] + + is_first_block = in_channels == out_channels == temb_channels + is_middle_block = in_channels != out_channels + add_self_attention = True if is_first_block else False + + self.has_cross_attention = True + self.attention_head_dim = attention_head_dim + + # in_channels, and out_channels for the block (k-unet) + k_in_channels = out_channels if is_first_block else 2 * out_channels + k_out_channels = in_channels + + num_layers = num_layers - 1 + + for i in range(num_layers): + in_channels = k_in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + if is_middle_block and (i == num_layers - 1): + conv_2d_out_channels = k_out_channels + else: + conv_2d_out_channels = None + + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + conv_2d_out_channels=conv_2d_out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=groups, + groups_out=groups_out, + dropout=dropout, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + attentions.append( + KAttentionBlock( + k_out_channels if (i == num_layers - 1) else out_channels, + k_out_channels // attention_head_dim + if (i == num_layers - 1) + else out_channels // attention_head_dim, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + temb_channels=temb_channels, + attention_bias=True, + add_self_attention=add_self_attention, + cross_attention_norm="layer_norm", + upcast_attention=upcast_attention, + ) + ) + + self.resnets = nn.ModuleList(resnets) + self.attentions = nn.ModuleList(attentions) + + if add_upsample: + self.upsamplers = nn.ModuleList([KUpsample2D()]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + res_hidden_states_tuple = res_hidden_states_tuple[-1] + if res_hidden_states_tuple is not None: + hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1) + + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +# can potentially later be renamed to `No-feed-forward` attention +class KAttentionBlock(nn.Module): + r""" + A basic Transformer block. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + attention_bias (`bool`, *optional*, defaults to `False`): + Configure if the attention layers should contain a bias parameter. + upcast_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to upcast the attention computation to `float32`. + temb_channels (`int`, *optional*, defaults to 768): + The number of channels in the token embedding. + add_self_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to add self-attention to the block. + cross_attention_norm (`str`, *optional*, defaults to `None`): + The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`. + group_size (`int`, *optional*, defaults to 32): + The number of groups to separate the channels into for group normalization. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout: float = 0.0, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + upcast_attention: bool = False, + temb_channels: int = 768, # for ada_group_norm + add_self_attention: bool = False, + cross_attention_norm: Optional[str] = None, + group_size: int = 32, + ): + super().__init__() + self.add_self_attention = add_self_attention + + # 1. Self-Attn + if add_self_attention: + self.norm1 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size)) + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=None, + cross_attention_norm=None, + ) + + # 2. Cross-Attn + self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size)) + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + cross_attention_norm=cross_attention_norm, + ) + + def _to_3d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor: + return hidden_states.permute(0, 2, 3, 1).reshape(hidden_states.shape[0], height * weight, -1) + + def _to_4d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor: + return hidden_states.permute(0, 2, 1).reshape(hidden_states.shape[0], -1, height, weight) + + def forward( + self, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + # TODO: mark emb as non-optional (self.norm2 requires it). + # requires assessing impact of change to positional param interface. + emb: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + # 1. Self-Attention + if self.add_self_attention: + norm_hidden_states = self.norm1(hidden_states, emb) + + height, weight = norm_hidden_states.shape[2:] + norm_hidden_states = self._to_3d(norm_hidden_states, height, weight) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + attn_output = self._to_4d(attn_output, height, weight) + + hidden_states = attn_output + hidden_states + + # 2. Cross-Attention/None + norm_hidden_states = self.norm2(hidden_states, emb) + + height, weight = norm_hidden_states.shape[2:] + norm_hidden_states = self._to_3d(norm_hidden_states, height, weight) + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask, + **cross_attention_kwargs, + ) + attn_output = self._to_4d(attn_output, height, weight) + + hidden_states = attn_output + hidden_states + + return hidden_states diff --git a/extensions-builtin/forge_space_idm_vton/src/unet_block_hacked_tryon.py b/extensions-builtin/forge_space_idm_vton/src/unet_block_hacked_tryon.py new file mode 100644 index 00000000..d46728a7 --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/src/unet_block_hacked_tryon.py @@ -0,0 +1,3522 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Any, Dict, Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import is_torch_version, logging +from diffusers.utils.torch_utils import apply_freeu +from diffusers.models.activations import get_activation +from diffusers.models.attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0 +from diffusers.models.dual_transformer_2d import DualTransformer2DModel +from diffusers.models.normalization import AdaGroupNorm +from diffusers.models.resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D +from src.transformerhacked_tryon import Transformer2DModel +from einops import rearrange + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def get_down_block( + down_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + temb_channels: int, + add_downsample: bool, + resnet_eps: float, + resnet_act_fn: str, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + downsample_padding: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + downsample_type: Optional[str] = None, + dropout: float = 0.0, +): + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warn( + f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type + if down_block_type == "DownBlock2D": + return DownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "ResnetDownsampleBlock2D": + return ResnetDownsampleBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + ) + elif down_block_type == "AttnDownBlock2D": + if add_downsample is False: + downsample_type = None + else: + downsample_type = downsample_type or "conv" # default to 'conv' + return AttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + downsample_type=downsample_type, + ) + elif down_block_type == "CrossAttnDownBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D") + return CrossAttnDownBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + elif down_block_type == "SimpleCrossAttnDownBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D") + return SimpleCrossAttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif down_block_type == "SkipDownBlock2D": + return SkipDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "AttnSkipDownBlock2D": + return AttnSkipDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "DownEncoderBlock2D": + return DownEncoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "AttnDownEncoderBlock2D": + return AttnDownEncoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "KDownBlock2D": + return KDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + ) + elif down_block_type == "KCrossAttnDownBlock2D": + return KCrossAttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + add_self_attention=True if not add_downsample else False, + ) + raise ValueError(f"{down_block_type} does not exist.") + + +def get_up_block( + up_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + add_upsample: bool, + resnet_eps: float, + resnet_act_fn: str, + resolution_idx: Optional[int] = None, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + upsample_type: Optional[str] = None, + dropout: float = 0.0, +) -> nn.Module: + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warn( + f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type + if up_block_type == "UpBlock2D": + return UpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "ResnetUpsampleBlock2D": + return ResnetUpsampleBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + ) + elif up_block_type == "CrossAttnUpBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D") + return CrossAttnUpBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + elif up_block_type == "SimpleCrossAttnUpBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D") + return SimpleCrossAttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif up_block_type == "AttnUpBlock2D": + if add_upsample is False: + upsample_type = None + else: + upsample_type = upsample_type or "conv" # default to 'conv' + + return AttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + upsample_type=upsample_type, + ) + elif up_block_type == "SkipUpBlock2D": + return SkipUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "AttnSkipUpBlock2D": + return AttnSkipUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "UpDecoderBlock2D": + return UpDecoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + temb_channels=temb_channels, + ) + elif up_block_type == "AttnUpDecoderBlock2D": + return AttnUpDecoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + temb_channels=temb_channels, + ) + elif up_block_type == "KUpBlock2D": + return KUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + ) + elif up_block_type == "KCrossAttnUpBlock2D": + return KCrossAttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + ) + + raise ValueError(f"{up_block_type} does not exist.") + + +class AutoencoderTinyBlock(nn.Module): + """ + Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU + blocks. + + Args: + in_channels (`int`): The number of input channels. + out_channels (`int`): The number of output channels. + act_fn (`str`): + ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`. + + Returns: + `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to + `out_channels`. + """ + + def __init__(self, in_channels: int, out_channels: int, act_fn: str): + super().__init__() + act_fn = get_activation(act_fn) + self.conv = nn.Sequential( + nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1), + act_fn, + nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1), + act_fn, + nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1), + ) + self.skip = ( + nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False) + if in_channels != out_channels + else nn.Identity() + ) + self.fuse = nn.ReLU() + + def forward(self, x: torch.FloatTensor) -> torch.FloatTensor: + return self.fuse(self.conv(x) + self.skip(x)) + + +class UNetMidBlock2D(nn.Module): + """ + A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks. + + Args: + in_channels (`int`): The number of input channels. + temb_channels (`int`): The number of temporal embedding channels. + dropout (`float`, *optional*, defaults to 0.0): The dropout rate. + num_layers (`int`, *optional*, defaults to 1): The number of residual blocks. + resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks. + resnet_time_scale_shift (`str`, *optional*, defaults to `default`): + The type of normalization to apply to the time embeddings. This can help to improve the performance of the + model on tasks with long-range temporal dependencies. + resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks. + resnet_groups (`int`, *optional*, defaults to 32): + The number of groups to use in the group normalization layers of the resnet blocks. + attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks. + resnet_pre_norm (`bool`, *optional*, defaults to `True`): + Whether to use pre-normalization for the resnet blocks. + add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks. + attention_head_dim (`int`, *optional*, defaults to 1): + Dimension of a single attention head. The number of attention heads is determined based on this value and + the number of input channels. + output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor. + + Returns: + `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size, + in_channels, height, width)`. + + """ + + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + attn_groups: Optional[int] = None, + resnet_pre_norm: bool = True, + add_attention: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + ): + super().__init__() + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + self.add_attention = add_attention + + if attn_groups is None: + attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}." + ) + attention_head_dim = in_channels + + for _ in range(num_layers): + if self.add_attention: + attentions.append( + Attention( + in_channels, + heads=in_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=attn_groups, + spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + else: + attentions.append(None) + + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor: + hidden_states = self.resnets[0](hidden_states, temb) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if attn is not None: + hidden_states = attn(hidden_states, temb=temb) + hidden_states = resnet(hidden_states, temb) + + return hidden_states + + +class UNetMidBlock2DCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + # support for variable transformer layers per block + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + for i in range(num_layers): + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + in_channels // num_attention_heads, + in_channels=in_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + in_channels // num_attention_heads, + in_channels=in_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + garment_features=None, + curr_garment_feat_idx=0, + ) -> torch.FloatTensor: + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states,curr_garment_feat_idx = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + garment_features=garment_features, + curr_garment_feat_idx=curr_garment_feat_idx, + ) + hidden_states=hidden_states[0] + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + else: + hidden_states,curr_garment_feat_idx = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + garment_features=garment_features, + curr_garment_feat_idx=curr_garment_feat_idx, + ) + hidden_states=hidden_states[0] + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + return hidden_states,curr_garment_feat_idx + + +class UNetMidBlock2DSimpleCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + + self.has_cross_attention = True + + self.attention_head_dim = attention_head_dim + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + self.num_heads = in_channels // self.attention_head_dim + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ] + attentions = [] + + for _ in range(num_layers): + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=in_channels, + cross_attention_dim=in_channels, + heads=self.num_heads, + dim_head=self.attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + # attn + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + # resnet + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class AttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + downsample_type: str = "conv", + ): + super().__init__() + resnets = [] + attentions = [] + self.downsample_type = downsample_type + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if downsample_type == "conv": + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + elif downsample_type == "resnet": + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + cross_attention_kwargs.update({"scale": lora_scale}) + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn(hidden_states, **cross_attention_kwargs) + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + if self.downsample_type == "resnet": + hidden_states = downsampler(hidden_states, temb=temb, scale=lora_scale) + else: + hidden_states = downsampler(hidden_states, scale=lora_scale) + + output_states += (hidden_states,) + + return hidden_states, output_states + + +class CrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + add_downsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + additional_residuals: Optional[torch.FloatTensor] = None, + garment_features=None, + curr_garment_feat_idx=0, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + blocks = list(zip(self.resnets, self.attentions)) + # print("len(self.resnets)") + # print(len(self.resnets)) + # print("len(self.attentions)") + # print(len(self.attentions)) + for i, (resnet, attn) in enumerate(blocks): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states,curr_garment_feat_idx = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + garment_features=garment_features, + curr_garment_feat_idx=curr_garment_feat_idx, + ) + hidden_states=hidden_states[0] + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states,curr_garment_feat_idx = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + garment_features=garment_features, + curr_garment_feat_idx=curr_garment_feat_idx, + ) + hidden_states=hidden_states[0] + + + # apply additional residuals to the output of the last pair of resnet and attention blocks + if i == len(blocks) - 1 and additional_residuals is not None: + hidden_states = hidden_states + additional_residuals + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale=lora_scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states,curr_garment_feat_idx + + +class DownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale=scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class DownEncoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor: + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=None, scale=scale) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale) + + return hidden_states + + +class AttnDownEncoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb=None, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale) + + return hidden_states + + +class AttnSkipDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = np.sqrt(2.0), + add_downsample: bool = True, + ): + super().__init__() + self.attentions = nn.ModuleList([]) + self.resnets = nn.ModuleList([]) + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + self.resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(in_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + self.attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=32, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + if add_downsample: + self.resnet_down = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + down=True, + kernel="fir", + ) + self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)]) + self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1)) + else: + self.resnet_down = None + self.downsamplers = None + self.skip_conv = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + skip_sample: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]: + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + output_states += (hidden_states,) + + if self.downsamplers is not None: + hidden_states = self.resnet_down(hidden_states, temb, scale=scale) + for downsampler in self.downsamplers: + skip_sample = downsampler(skip_sample) + + hidden_states = self.skip_conv(skip_sample) + hidden_states + + output_states += (hidden_states,) + + return hidden_states, output_states, skip_sample + + +class SkipDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + output_scale_factor: float = np.sqrt(2.0), + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + self.resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(in_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + if add_downsample: + self.resnet_down = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + down=True, + kernel="fir", + ) + self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)]) + self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1)) + else: + self.resnet_down = None + self.downsamplers = None + self.skip_conv = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + skip_sample: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]: + output_states = () + + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb, scale) + output_states += (hidden_states,) + + if self.downsamplers is not None: + hidden_states = self.resnet_down(hidden_states, temb, scale) + for downsampler in self.downsamplers: + skip_sample = downsampler(skip_sample) + + hidden_states = self.skip_conv(skip_sample) + hidden_states + + output_states += (hidden_states,) + + return hidden_states, output_states, skip_sample + + +class ResnetDownsampleBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + skip_time_act: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, temb, scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class SimpleCrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + + self.has_cross_attention = True + + resnets = [] + attentions = [] + + self.attention_head_dim = attention_head_dim + self.num_heads = out_channels // self.attention_head_dim + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=out_channels, + cross_attention_dim=out_channels, + heads=self.num_heads, + dim_head=attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, temb, scale=lora_scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class KDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: int = 32, + add_downsample: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + temb_channels=temb_channels, + groups=groups, + groups_out=groups_out, + eps=resnet_eps, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + # YiYi's comments- might be able to use FirDownsample2D, look into details later + self.downsamplers = nn.ModuleList([KDownsample2D()]) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale) + + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + return hidden_states, output_states + + +class KCrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + cross_attention_dim: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_group_size: int = 32, + add_downsample: bool = True, + attention_head_dim: int = 64, + add_self_attention: bool = False, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + temb_channels=temb_channels, + groups=groups, + groups_out=groups_out, + eps=resnet_eps, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + attentions.append( + KAttentionBlock( + out_channels, + out_channels // attention_head_dim, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + temb_channels=temb_channels, + attention_bias=True, + add_self_attention=add_self_attention, + cross_attention_norm="layer_norm", + group_size=resnet_group_size, + ) + ) + + self.resnets = nn.ModuleList(resnets) + self.attentions = nn.ModuleList(attentions) + + if add_downsample: + self.downsamplers = nn.ModuleList([KDownsample2D()]) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + + if self.downsamplers is None: + output_states += (None,) + else: + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + return hidden_states, output_states + + +class AttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: int = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + upsample_type: str = "conv", + ): + super().__init__() + resnets = [] + attentions = [] + + self.upsample_type = upsample_type + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if upsample_type == "conv": + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + elif upsample_type == "resnet": + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + if self.upsample_type == "resnet": + hidden_states = upsampler(hidden_states, temb=temb, scale=scale) + else: + hidden_states = upsampler(hidden_states, scale=scale) + + return hidden_states + + +class CrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + garment_features=None, + curr_garment_feat_idx=0, + ) -> torch.FloatTensor: + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + + for resnet, attn in zip(self.resnets, self.attentions): + + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + # print(hidden_states.shape) + # print(encoder_hidden_states.shape) + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states,curr_garment_feat_idx = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + garment_features=garment_features, + curr_garment_feat_idx=curr_garment_feat_idx, + ) + hidden_states=hidden_states[0] + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states,curr_garment_feat_idx = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + garment_features=garment_features, + curr_garment_feat_idx=curr_garment_feat_idx, + ) + hidden_states=hidden_states[0] + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale) + + + + return hidden_states,curr_garment_feat_idx + + +class UpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size, scale=scale) + + return hidden_states + + + +class UpDecoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + temb_channels: Optional[int] = None, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + input_channels = in_channels if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> torch.FloatTensor: + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +class AttnUpDecoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + temb_channels: Optional[int] = None, + ): + super().__init__() + resnets = [] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + input_channels = in_channels if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None, + spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb=temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, temb=temb, **cross_attention_kwargs) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, scale=scale) + + return hidden_states + + +class AttnSkipUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = np.sqrt(2.0), + add_upsample: bool = True, + ): + super().__init__() + self.attentions = nn.ModuleList([]) + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + self.resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(resnet_in_channels + res_skip_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + self.attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=32, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels) + if add_upsample: + self.resnet_up = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + up=True, + kernel="fir", + ) + self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + self.skip_norm = torch.nn.GroupNorm( + num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True + ) + self.act = nn.SiLU() + else: + self.resnet_up = None + self.skip_conv = None + self.skip_norm = None + self.act = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + skip_sample=None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, torch.FloatTensor]: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + + cross_attention_kwargs = {"scale": scale} + hidden_states = self.attentions[0](hidden_states, **cross_attention_kwargs) + + if skip_sample is not None: + skip_sample = self.upsampler(skip_sample) + else: + skip_sample = 0 + + if self.resnet_up is not None: + skip_sample_states = self.skip_norm(hidden_states) + skip_sample_states = self.act(skip_sample_states) + skip_sample_states = self.skip_conv(skip_sample_states) + + skip_sample = skip_sample + skip_sample_states + + hidden_states = self.resnet_up(hidden_states, temb, scale=scale) + + return hidden_states, skip_sample + + +class SkipUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + output_scale_factor: float = np.sqrt(2.0), + add_upsample: bool = True, + upsample_padding: int = 1, + ): + super().__init__() + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + self.resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min((resnet_in_channels + res_skip_channels) // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels) + if add_upsample: + self.resnet_up = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + up=True, + kernel="fir", + ) + self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + self.skip_norm = torch.nn.GroupNorm( + num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True + ) + self.act = nn.SiLU() + else: + self.resnet_up = None + self.skip_conv = None + self.skip_norm = None + self.act = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + skip_sample=None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, torch.FloatTensor]: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + + if skip_sample is not None: + skip_sample = self.upsampler(skip_sample) + else: + skip_sample = 0 + + if self.resnet_up is not None: + skip_sample_states = self.skip_norm(hidden_states) + skip_sample_states = self.act(skip_sample_states) + skip_sample_states = self.skip_conv(skip_sample_states) + + skip_sample = skip_sample + skip_sample_states + + hidden_states = self.resnet_up(hidden_states, temb, scale=scale) + + return hidden_states, skip_sample + + +class ResnetUpsampleBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + skip_time_act: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, temb, scale=scale) + + return hidden_states + + +class SimpleCrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.attention_head_dim = attention_head_dim + + self.num_heads = out_channels // self.attention_head_dim + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=out_channels, + cross_attention_dim=out_channels, + heads=self.num_heads, + dim_head=self.attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + for resnet, attn in zip(self.resnets, self.attentions): + # resnet + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class KUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + resolution_idx: int, + dropout: float = 0.0, + num_layers: int = 5, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: Optional[int] = 32, + add_upsample: bool = True, + ): + super().__init__() + resnets = [] + k_in_channels = 2 * out_channels + k_out_channels = in_channels + num_layers = num_layers - 1 + + for i in range(num_layers): + in_channels = k_in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=k_out_channels if (i == num_layers - 1) else out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=groups, + groups_out=groups_out, + dropout=dropout, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([KUpsample2D()]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + res_hidden_states_tuple = res_hidden_states_tuple[-1] + if res_hidden_states_tuple is not None: + hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1) + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +class KCrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + resolution_idx: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: int = 32, + attention_head_dim: int = 1, # attention dim_head + cross_attention_dim: int = 768, + add_upsample: bool = True, + upcast_attention: bool = False, + ): + super().__init__() + resnets = [] + attentions = [] + + is_first_block = in_channels == out_channels == temb_channels + is_middle_block = in_channels != out_channels + add_self_attention = True if is_first_block else False + + self.has_cross_attention = True + self.attention_head_dim = attention_head_dim + + # in_channels, and out_channels for the block (k-unet) + k_in_channels = out_channels if is_first_block else 2 * out_channels + k_out_channels = in_channels + + num_layers = num_layers - 1 + + for i in range(num_layers): + in_channels = k_in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + if is_middle_block and (i == num_layers - 1): + conv_2d_out_channels = k_out_channels + else: + conv_2d_out_channels = None + + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + conv_2d_out_channels=conv_2d_out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=groups, + groups_out=groups_out, + dropout=dropout, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + attentions.append( + KAttentionBlock( + k_out_channels if (i == num_layers - 1) else out_channels, + k_out_channels // attention_head_dim + if (i == num_layers - 1) + else out_channels // attention_head_dim, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + temb_channels=temb_channels, + attention_bias=True, + add_self_attention=add_self_attention, + cross_attention_norm="layer_norm", + upcast_attention=upcast_attention, + ) + ) + + self.resnets = nn.ModuleList(resnets) + self.attentions = nn.ModuleList(attentions) + + if add_upsample: + self.upsamplers = nn.ModuleList([KUpsample2D()]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + res_hidden_states_tuple = res_hidden_states_tuple[-1] + if res_hidden_states_tuple is not None: + hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1) + + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +# can potentially later be renamed to `No-feed-forward` attention +class KAttentionBlock(nn.Module): + r""" + A basic Transformer block. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + attention_bias (`bool`, *optional*, defaults to `False`): + Configure if the attention layers should contain a bias parameter. + upcast_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to upcast the attention computation to `float32`. + temb_channels (`int`, *optional*, defaults to 768): + The number of channels in the token embedding. + add_self_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to add self-attention to the block. + cross_attention_norm (`str`, *optional*, defaults to `None`): + The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`. + group_size (`int`, *optional*, defaults to 32): + The number of groups to separate the channels into for group normalization. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout: float = 0.0, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + upcast_attention: bool = False, + temb_channels: int = 768, # for ada_group_norm + add_self_attention: bool = False, + cross_attention_norm: Optional[str] = None, + group_size: int = 32, + ): + super().__init__() + self.add_self_attention = add_self_attention + + # 1. Self-Attn + if add_self_attention: + self.norm1 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size)) + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=None, + cross_attention_norm=None, + ) + + # 2. Cross-Attn + self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size)) + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + cross_attention_norm=cross_attention_norm, + ) + + def _to_3d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor: + return hidden_states.permute(0, 2, 3, 1).reshape(hidden_states.shape[0], height * weight, -1) + + def _to_4d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor: + return hidden_states.permute(0, 2, 1).reshape(hidden_states.shape[0], -1, height, weight) + + def forward( + self, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + # TODO: mark emb as non-optional (self.norm2 requires it). + # requires assessing impact of change to positional param interface. + emb: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + # 1. Self-Attention + if self.add_self_attention: + norm_hidden_states = self.norm1(hidden_states, emb) + + height, weight = norm_hidden_states.shape[2:] + norm_hidden_states = self._to_3d(norm_hidden_states, height, weight) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + attn_output = self._to_4d(attn_output, height, weight) + + hidden_states = attn_output + hidden_states + + # 2. Cross-Attention/None + norm_hidden_states = self.norm2(hidden_states, emb) + + height, weight = norm_hidden_states.shape[2:] + norm_hidden_states = self._to_3d(norm_hidden_states, height, weight) + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask, + **cross_attention_kwargs, + ) + attn_output = self._to_4d(attn_output, height, weight) + + hidden_states = attn_output + hidden_states + + return hidden_states diff --git a/extensions-builtin/forge_space_idm_vton/src/unet_hacked_garmnet.py b/extensions-builtin/forge_space_idm_vton/src/unet_hacked_garmnet.py new file mode 100644 index 00000000..53b1f857 --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/src/unet_hacked_garmnet.py @@ -0,0 +1,1284 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, List, Optional, Tuple, Union + +import torch +import torch.nn as nn +import torch.utils.checkpoint + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import UNet2DConditionLoadersMixin +from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers +from diffusers.models.activations import get_activation +from diffusers.models.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + Attention, + AttentionProcessor, + AttnAddedKVProcessor, + AttnProcessor, +) +from einops import rearrange + +from diffusers.models.embeddings import ( + GaussianFourierProjection, + ImageHintTimeEmbedding, + ImageProjection, + ImageTimeEmbedding, + PositionNet, + TextImageProjection, + TextImageTimeEmbedding, + TextTimeEmbedding, + TimestepEmbedding, + Timesteps, +) +from diffusers.models.modeling_utils import ModelMixin +from src.unet_block_hacked_garmnet import ( + UNetMidBlock2D, + UNetMidBlock2DCrossAttn, + UNetMidBlock2DSimpleCrossAttn, + get_down_block, + get_up_block, +) +from diffusers.models.resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D +from diffusers.models.transformer_2d import Transformer2DModel + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def zero_module(module): + for p in module.parameters(): + nn.init.zeros_(p) + return module + +@dataclass +class UNet2DConditionOutput(BaseOutput): + """ + The output of [`UNet2DConditionModel`]. + + Args: + sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): + The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model. + """ + + sample: torch.FloatTensor = None + + +class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin): + r""" + A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample + shaped output. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`): + Height and width of input/output sample. + in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample. + out_channels (`int`, *optional*, defaults to 4): Number of channels in the output. + center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample. + flip_sin_to_cos (`bool`, *optional*, defaults to `False`): + Whether to flip the sin to cos in the time embedding. + freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding. + down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`): + The tuple of downsample blocks to use. + mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`): + Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or + `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`): + The tuple of upsample blocks to use. + only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`): + Whether to include self-attention in the basic transformer blocks, see + [`~models.attention.BasicTransformerBlock`]. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`): + The tuple of output channels for each block. + layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block. + downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution. + mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization. + If `None`, normalization and activation layers is skipped in post-processing. + norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization. + cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280): + The dimension of the cross attention features. + transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling + blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + encoder_hid_dim (`int`, *optional*, defaults to None): + If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim` + dimension to `cross_attention_dim`. + encoder_hid_dim_type (`str`, *optional*, defaults to `None`): + If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text + embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`. + attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads. + num_attention_heads (`int`, *optional*): + The number of attention heads. If not defined, defaults to `attention_head_dim` + resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config + for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`. + class_embed_type (`str`, *optional*, defaults to `None`): + The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`, + `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`. + addition_embed_type (`str`, *optional*, defaults to `None`): + Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or + "text". "text" will use the `TextTimeEmbedding` layer. + addition_time_embed_dim: (`int`, *optional*, defaults to `None`): + Dimension for the timestep embeddings. + num_class_embeds (`int`, *optional*, defaults to `None`): + Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing + class conditioning with `class_embed_type` equal to `None`. + time_embedding_type (`str`, *optional*, defaults to `positional`): + The type of position embedding to use for timesteps. Choose from `positional` or `fourier`. + time_embedding_dim (`int`, *optional*, defaults to `None`): + An optional override for the dimension of the projected time embedding. + time_embedding_act_fn (`str`, *optional*, defaults to `None`): + Optional activation function to use only once on the time embeddings before they are passed to the rest of + the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`. + timestep_post_act (`str`, *optional*, defaults to `None`): + The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`. + time_cond_proj_dim (`int`, *optional*, defaults to `None`): + The dimension of `cond_proj` layer in the timestep embedding. + conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`, + *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`, + *optional*): The dimension of the `class_labels` input when + `class_embed_type="projection"`. Required when `class_embed_type="projection"`. + class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time + embeddings with the class embeddings. + mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`): + Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If + `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the + `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False` + otherwise. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + sample_size: Optional[int] = None, + in_channels: int = 4, + out_channels: int = 4, + center_input_sample: bool = False, + flip_sin_to_cos: bool = True, + freq_shift: int = 0, + down_block_types: Tuple[str] = ( + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "DownBlock2D", + ), + mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn", + up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"), + only_cross_attention: Union[bool, Tuple[bool]] = False, + block_out_channels: Tuple[int] = (320, 640, 1280, 1280), + layers_per_block: Union[int, Tuple[int]] = 2, + downsample_padding: int = 1, + mid_block_scale_factor: float = 1, + dropout: float = 0.0, + act_fn: str = "silu", + norm_num_groups: Optional[int] = 32, + norm_eps: float = 1e-5, + cross_attention_dim: Union[int, Tuple[int]] = 1280, + transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1, + reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None, + encoder_hid_dim: Optional[int] = None, + encoder_hid_dim_type: Optional[str] = None, + attention_head_dim: Union[int, Tuple[int]] = 8, + num_attention_heads: Optional[Union[int, Tuple[int]]] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + class_embed_type: Optional[str] = None, + addition_embed_type: Optional[str] = None, + addition_time_embed_dim: Optional[int] = None, + num_class_embeds: Optional[int] = None, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: int = 1.0, + time_embedding_type: str = "positional", + time_embedding_dim: Optional[int] = None, + time_embedding_act_fn: Optional[str] = None, + timestep_post_act: Optional[str] = None, + time_cond_proj_dim: Optional[int] = None, + conv_in_kernel: int = 3, + conv_out_kernel: int = 3, + projection_class_embeddings_input_dim: Optional[int] = None, + attention_type: str = "default", + class_embeddings_concat: bool = False, + mid_block_only_cross_attention: Optional[bool] = None, + cross_attention_norm: Optional[str] = None, + addition_embed_type_num_heads=64, + ): + super().__init__() + + self.sample_size = sample_size + + if num_attention_heads is not None: + raise ValueError( + "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19." + ) + + # If `num_attention_heads` is not defined (which is the case for most models) + # it will default to `attention_head_dim`. This looks weird upon first reading it and it is. + # The reason for this behavior is to correct for incorrectly named variables that were introduced + # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 + # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking + # which is why we correct for the naming here. + num_attention_heads = num_attention_heads or attention_head_dim + + # Check inputs + if len(down_block_types) != len(up_block_types): + raise ValueError( + f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}." + ) + + if len(block_out_channels) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}." + ) + + if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}." + ) + if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None: + for layer_number_per_block in transformer_layers_per_block: + if isinstance(layer_number_per_block, list): + raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.") + + # input + conv_in_padding = (conv_in_kernel - 1) // 2 + self.conv_in = nn.Conv2d( + in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding + ) + + # time + if time_embedding_type == "fourier": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 2 + if time_embed_dim % 2 != 0: + raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.") + self.time_proj = GaussianFourierProjection( + time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos + ) + timestep_input_dim = time_embed_dim + elif time_embedding_type == "positional": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 4 + + self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift) + timestep_input_dim = block_out_channels[0] + else: + raise ValueError( + f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`." + ) + + self.time_embedding = TimestepEmbedding( + timestep_input_dim, + time_embed_dim, + act_fn=act_fn, + post_act_fn=timestep_post_act, + cond_proj_dim=time_cond_proj_dim, + ) + + if encoder_hid_dim_type is None and encoder_hid_dim is not None: + encoder_hid_dim_type = "text_proj" + self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type) + logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.") + + if encoder_hid_dim is None and encoder_hid_dim_type is not None: + raise ValueError( + f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}." + ) + + if encoder_hid_dim_type == "text_proj": + self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim) + elif encoder_hid_dim_type == "text_image_proj": + # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)` + self.encoder_hid_proj = TextImageProjection( + text_embed_dim=encoder_hid_dim, + image_embed_dim=cross_attention_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 + self.encoder_hid_proj = ImageProjection( + image_embed_dim=encoder_hid_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type is not None: + raise ValueError( + f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'." + ) + else: + self.encoder_hid_proj = None + + # class embedding + if class_embed_type is None and num_class_embeds is not None: + self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim) + elif class_embed_type == "timestep": + self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn) + elif class_embed_type == "identity": + self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim) + elif class_embed_type == "projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set" + ) + # The projection `class_embed_type` is the same as the timestep `class_embed_type` except + # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings + # 2. it projects from an arbitrary input dimension. + # + # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations. + # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings. + # As a result, `TimestepEmbedding` can be passed arbitrary vectors. + self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif class_embed_type == "simple_projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set" + ) + self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim) + else: + self.class_embedding = None + + if addition_embed_type == "text": + if encoder_hid_dim is not None: + text_time_embedding_from_dim = encoder_hid_dim + else: + text_time_embedding_from_dim = cross_attention_dim + + self.add_embedding = TextTimeEmbedding( + text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads + ) + elif addition_embed_type == "text_image": + # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)` + self.add_embedding = TextImageTimeEmbedding( + text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim + ) + elif addition_embed_type == "text_time": + self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift) + self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif addition_embed_type == "image": + # Kandinsky 2.2 + self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type == "image_hint": + # Kandinsky 2.2 ControlNet + self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type is not None: + raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.") + + if time_embedding_act_fn is None: + self.time_embed_act = None + else: + self.time_embed_act = get_activation(time_embedding_act_fn) + + self.down_blocks = nn.ModuleList([]) + self.up_blocks = nn.ModuleList([]) + + if isinstance(only_cross_attention, bool): + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = only_cross_attention + + only_cross_attention = [only_cross_attention] * len(down_block_types) + + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = False + + if isinstance(num_attention_heads, int): + num_attention_heads = (num_attention_heads,) * len(down_block_types) + + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * len(down_block_types) + + if isinstance(cross_attention_dim, int): + cross_attention_dim = (cross_attention_dim,) * len(down_block_types) + + if isinstance(layers_per_block, int): + layers_per_block = [layers_per_block] * len(down_block_types) + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types) + if class_embeddings_concat: + # The time embeddings are concatenated with the class embeddings. The dimension of the + # time embeddings passed to the down, middle, and up blocks is twice the dimension of the + # regular time embeddings + blocks_time_embed_dim = time_embed_dim * 2 + else: + blocks_time_embed_dim = time_embed_dim + + # down + output_channel = block_out_channels[0] + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + + down_block = get_down_block( + down_block_type, + num_layers=layers_per_block[i], + transformer_layers_per_block=transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + temb_channels=blocks_time_embed_dim, + add_downsample=not is_final_block, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim[i], + num_attention_heads=num_attention_heads[i], + downsample_padding=downsample_padding, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.down_blocks.append(down_block) + + # mid + if mid_block_type == "UNetMidBlock2DCrossAttn": + self.mid_block = UNetMidBlock2DCrossAttn( + transformer_layers_per_block=transformer_layers_per_block[-1], + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_time_scale_shift=resnet_time_scale_shift, + cross_attention_dim=cross_attention_dim[-1], + num_attention_heads=num_attention_heads[-1], + resnet_groups=norm_num_groups, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn": + self.mid_block = UNetMidBlock2DSimpleCrossAttn( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + cross_attention_dim=cross_attention_dim[-1], + attention_head_dim=attention_head_dim[-1], + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + only_cross_attention=mid_block_only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif mid_block_type == "UNetMidBlock2D": + self.mid_block = UNetMidBlock2D( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + num_layers=0, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + add_attention=False, + ) + elif mid_block_type is None: + self.mid_block = None + else: + raise ValueError(f"unknown mid_block_type : {mid_block_type}") + + # count how many layers upsample the images + self.num_upsamplers = 0 + + # up + reversed_block_out_channels = list(reversed(block_out_channels)) + reversed_num_attention_heads = list(reversed(num_attention_heads)) + reversed_layers_per_block = list(reversed(layers_per_block)) + reversed_cross_attention_dim = list(reversed(cross_attention_dim)) + reversed_transformer_layers_per_block = ( + list(reversed(transformer_layers_per_block)) + if reverse_transformer_layers_per_block is None + else reverse_transformer_layers_per_block + ) + only_cross_attention = list(reversed(only_cross_attention)) + + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(up_block_types): + is_final_block = i == len(block_out_channels) - 1 + + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)] + + # add upsample block for all BUT final layer + if not is_final_block: + add_upsample = True + self.num_upsamplers += 1 + else: + add_upsample = False + up_block = get_up_block( + up_block_type, + num_layers=reversed_layers_per_block[i] + 1, + transformer_layers_per_block=reversed_transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + prev_output_channel=prev_output_channel, + temb_channels=blocks_time_embed_dim, + add_upsample=add_upsample, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resolution_idx=i, + resnet_groups=norm_num_groups, + cross_attention_dim=reversed_cross_attention_dim[i], + num_attention_heads=reversed_num_attention_heads[i], + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + + + + # encode_output_chs = [ + # # 320, + # # 320, + # # 320, + # 1280, + # 1280, + # 1280, + # 1280, + # 640, + # 640 + # ] + + # encode_output_chs2 = [ + # # 320, + # # 320, + # # 320, + # 1280, + # 1280, + # 640, + # 640, + # 640, + # 320 + # ] + + # encode_num_head_chs3 = [ + # # 5, + # # 5, + # # 10, + # 20, + # 20, + # 20, + # 10, + # 10, + # 10 + # ] + + + # encode_num_layers_chs4 = [ + # # 1, + # # 1, + # # 2, + # 10, + # 10, + # 10, + # 2, + # 2, + # 2 + # ] + + + # self.warp_blks = nn.ModuleList([]) + # self.warp_zeros = nn.ModuleList([]) + + # for in_ch, cont_ch,num_head,num_layers in zip(encode_output_chs, encode_output_chs2,encode_num_head_chs3,encode_num_layers_chs4): + # # dim_head = in_ch // self.num_heads + # # dim_head = dim_head // dim_head_denorm + + # self.warp_blks.append(Transformer2DModel( + # num_attention_heads=num_head, + # attention_head_dim=64, + # in_channels=in_ch, + # num_layers = num_layers, + # cross_attention_dim = cont_ch, + # )) + + # self.warp_zeros.append(zero_module(nn.Conv2d(in_ch, in_ch, 1, padding=0))) + + + + # out + if norm_num_groups is not None: + self.conv_norm_out = nn.GroupNorm( + num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps + ) + + self.conv_act = get_activation(act_fn) + + else: + self.conv_norm_out = None + self.conv_act = None + + conv_out_padding = (conv_out_kernel - 1) // 2 + self.conv_out = nn.Conv2d( + block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding + ) + + if attention_type in ["gated", "gated-text-image"]: + positive_len = 768 + if isinstance(cross_attention_dim, int): + positive_len = cross_attention_dim + elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list): + positive_len = cross_attention_dim[0] + + feature_type = "text-only" if attention_type == "gated" else "text-image" + self.position_net = PositionNet( + positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type + ) + + + + + @property + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True) + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + def set_attn_processor( + self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False + ): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor, _remove_lora=True) + + def set_attention_slice(self, slice_size): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module splits the input tensor in slices to compute attention in + several steps. This is useful for saving some memory in exchange for a small decrease in speed. + + Args: + slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`): + When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If + `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is + provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` + must be a multiple of `slice_size`. + """ + sliceable_head_dims = [] + + def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module): + if hasattr(module, "set_attention_slice"): + sliceable_head_dims.append(module.sliceable_head_dim) + + for child in module.children(): + fn_recursive_retrieve_sliceable_dims(child) + + # retrieve number of attention layers + for module in self.children(): + fn_recursive_retrieve_sliceable_dims(module) + + num_sliceable_layers = len(sliceable_head_dims) + + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = [dim // 2 for dim in sliceable_head_dims] + elif slice_size == "max": + # make smallest slice possible + slice_size = num_sliceable_layers * [1] + + slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size + + if len(slice_size) != len(sliceable_head_dims): + raise ValueError( + f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different" + f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}." + ) + + for i in range(len(slice_size)): + size = slice_size[i] + dim = sliceable_head_dims[i] + if size is not None and size > dim: + raise ValueError(f"size {size} has to be smaller or equal to {dim}.") + + # Recursively walk through all the children. + # Any children which exposes the set_attention_slice method + # gets the message + def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]): + if hasattr(module, "set_attention_slice"): + module.set_attention_slice(slice_size.pop()) + + for child in module.children(): + fn_recursive_set_attention_slice(child, slice_size) + + reversed_slice_size = list(reversed(slice_size)) + for module in self.children(): + fn_recursive_set_attention_slice(module, reversed_slice_size) + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def enable_freeu(self, s1, s2, b1, b2): + r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497. + + The suffixes after the scaling factors represent the stage blocks where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that + are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + for i, upsample_block in enumerate(self.up_blocks): + setattr(upsample_block, "s1", s1) + setattr(upsample_block, "s2", s2) + setattr(upsample_block, "b1", b1) + setattr(upsample_block, "b2", b2) + + def disable_freeu(self): + """Disables the FreeU mechanism.""" + freeu_keys = {"s1", "s2", "b1", "b2"} + for i, upsample_block in enumerate(self.up_blocks): + for k in freeu_keys: + if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None: + setattr(upsample_block, k, None) + + def fuse_qkv_projections(self): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is ๐Ÿงช experimental. + + + """ + self.original_attn_processors = None + + for _, attn_processor in self.attn_processors.items(): + if "Added" in str(attn_processor.__class__.__name__): + raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.") + + self.original_attn_processors = self.attn_processors + + for module in self.modules(): + if isinstance(module, Attention): + module.fuse_projections(fuse=True) + + def unfuse_qkv_projections(self): + """Disables the fused QKV projection if enabled. + + + + This API is ๐Ÿงช experimental. + + + + """ + if self.original_attn_processors is not None: + self.set_attn_processor(self.original_attn_processors) + + def forward( + self, + sample: torch.FloatTensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + class_labels: Optional[torch.Tensor] = None, + timestep_cond: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + mid_block_additional_residual: Optional[torch.Tensor] = None, + down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + ) -> Union[UNet2DConditionOutput, Tuple]: + r""" + The [`UNet2DConditionModel`] forward method. + + Args: + sample (`torch.FloatTensor`): + The noisy input tensor with the following shape `(batch, channel, height, width)`. + timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input. + encoder_hidden_states (`torch.FloatTensor`): + The encoder hidden states with shape `(batch, sequence_length, feature_dim)`. + class_labels (`torch.Tensor`, *optional*, defaults to `None`): + Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings. + timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`): + Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed + through the `self.time_embedding` layer to obtain the timestep embeddings. + attention_mask (`torch.Tensor`, *optional*, defaults to `None`): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*): + A tuple of tensors that if specified are added to the residuals of down unet blocks. + mid_block_additional_residual: (`torch.Tensor`, *optional*): + A tensor that if specified is added to the residual of the middle unet block. + encoder_attention_mask (`torch.Tensor`): + A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If + `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias, + which adds large negative values to the attention scores corresponding to "discard" tokens. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttnProcessor`]. + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added to UNet long skip connections from down blocks to up blocks for + example from ControlNet side model(s) + mid_block_additional_residual (`torch.Tensor`, *optional*): + additional residual to be added to UNet mid block output, for example from ControlNet side model + down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s) + + Returns: + [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`: + If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise + a `tuple` is returned where the first element is the sample tensor. + """ + # By default samples have to be AT least a multiple of the overall upsampling factor. + # The overall upsampling factor is equal to 2 ** (# num of upsampling layers). + # However, the upsampling interpolation output size can be forced to fit any upsampling size + # on the fly if necessary. + default_overall_up_factor = 2**self.num_upsamplers + + # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` + forward_upsample_size = False + upsample_size = None + + for dim in sample.shape[-2:]: + if dim % default_overall_up_factor != 0: + # Forward upsample size to force interpolation output size. + forward_upsample_size = True + break + + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None: + encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # 0. center input if necessary + if self.config.center_input_sample: + sample = 2 * sample - 1.0 + + # 1. time + timesteps = timestep + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = sample.device.type == "mps" + if isinstance(timestep, float): + dtype = torch.float32 if is_mps else torch.float64 + else: + dtype = torch.int32 if is_mps else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + t_emb = self.time_proj(timesteps) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=sample.dtype) + + emb = self.time_embedding(t_emb, timestep_cond) + aug_emb = None + + if self.class_embedding is not None: + if class_labels is None: + raise ValueError("class_labels should be provided when num_class_embeds > 0") + + if self.config.class_embed_type == "timestep": + class_labels = self.time_proj(class_labels) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # there might be better ways to encapsulate this. + class_labels = class_labels.to(dtype=sample.dtype) + + class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype) + + if self.config.class_embeddings_concat: + emb = torch.cat([emb, class_emb], dim=-1) + else: + emb = emb + class_emb + + if self.config.addition_embed_type == "text": + aug_emb = self.add_embedding(encoder_hidden_states) + elif self.config.addition_embed_type == "text_image": + # Kandinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + + image_embs = added_cond_kwargs.get("image_embeds") + text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states) + aug_emb = self.add_embedding(text_embs, image_embs) + elif self.config.addition_embed_type == "text_time": + # SDXL - style + if "text_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`" + ) + text_embeds = added_cond_kwargs.get("text_embeds") + if "time_ids" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`" + ) + time_ids = added_cond_kwargs.get("time_ids") + time_embeds = self.add_time_proj(time_ids.flatten()) + time_embeds = time_embeds.reshape((text_embeds.shape[0], -1)) + add_embeds = torch.concat([text_embeds, time_embeds], dim=-1) + add_embeds = add_embeds.to(emb.dtype) + aug_emb = self.add_embedding(add_embeds) + elif self.config.addition_embed_type == "image": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + aug_emb = self.add_embedding(image_embs) + elif self.config.addition_embed_type == "image_hint": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + hint = added_cond_kwargs.get("hint") + aug_emb, hint = self.add_embedding(image_embs, hint) + sample = torch.cat([sample, hint], dim=1) + + emb = emb + aug_emb if aug_emb is not None else emb + + if self.time_embed_act is not None: + emb = self.time_embed_act(emb) + + if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj": + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj": + # Kadinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj": + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + image_embeds = self.encoder_hid_proj(image_embeds).to(encoder_hidden_states.dtype) + encoder_hidden_states = torch.cat([encoder_hidden_states, image_embeds], dim=1) + + # 2. pre-process + sample = self.conv_in(sample) + garment_features=[] + + # 2.5 GLIGEN position net + if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None: + cross_attention_kwargs = cross_attention_kwargs.copy() + gligen_args = cross_attention_kwargs.pop("gligen") + cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)} + + + # 3. down + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + if USE_PEFT_BACKEND: + # weight the lora layers by setting `lora_scale` for each PEFT layer + scale_lora_layers(self, lora_scale) + + is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None + # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets + is_adapter = down_intrablock_additional_residuals is not None + # maintain backward compatibility for legacy usage, where + # T2I-Adapter and ControlNet both use down_block_additional_residuals arg + # but can only use one or the other + if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None: + deprecate( + "T2I should not use down_block_additional_residuals", + "1.3.0", + "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \ + and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \ + for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ", + standard_warn=False, + ) + down_intrablock_additional_residuals = down_block_additional_residuals + is_adapter = True + + down_block_res_samples = (sample,) + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + # For t2i-adapter CrossAttnDownBlock2D + additional_residuals = {} + if is_adapter and len(down_intrablock_additional_residuals) > 0: + additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0) + + sample, res_samples,out_garment_feat = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + **additional_residuals, + ) + garment_features += out_garment_feat + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb, scale=lora_scale) + if is_adapter and len(down_intrablock_additional_residuals) > 0: + sample += down_intrablock_additional_residuals.pop(0) + + down_block_res_samples += res_samples + + + if is_controlnet: + new_down_block_res_samples = () + + for down_block_res_sample, down_block_additional_residual in zip( + down_block_res_samples, down_block_additional_residuals + ): + down_block_res_sample = down_block_res_sample + down_block_additional_residual + new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = new_down_block_res_samples + + # 4. mid + if self.mid_block is not None: + if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention: + sample,out_garment_feat = self.mid_block( + sample, + emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + garment_features += out_garment_feat + + else: + sample = self.mid_block(sample, emb) + + # To support T2I-Adapter-XL + if ( + is_adapter + and len(down_intrablock_additional_residuals) > 0 + and sample.shape == down_intrablock_additional_residuals[0].shape + ): + sample += down_intrablock_additional_residuals.pop(0) + + if is_controlnet: + sample = sample + mid_block_additional_residual + + + + # 5. up + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] + + # if we have not reached the final block and need to forward the + # upsample size, we do it here + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: + sample,out_garment_feat = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + upsample_size=upsample_size, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + ) + garment_features += out_garment_feat + + + if not return_dict: + return (sample,),garment_features + + return UNet2DConditionOutput(sample=sample),garment_features diff --git a/extensions-builtin/forge_space_idm_vton/src/unet_hacked_tryon.py b/extensions-builtin/forge_space_idm_vton/src/unet_hacked_tryon.py new file mode 100644 index 00000000..c6c795ce --- /dev/null +++ b/extensions-builtin/forge_space_idm_vton/src/unet_hacked_tryon.py @@ -0,0 +1,1395 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, List, Optional, Tuple, Union + +import torch +import torch.nn as nn +import torch.utils.checkpoint + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import UNet2DConditionLoadersMixin +from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers +from diffusers.models.activations import get_activation +from diffusers.models.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + Attention, + AttentionProcessor, + AttnAddedKVProcessor, + AttnProcessor, +) +from einops import rearrange + +from diffusers.models.embeddings import ( + GaussianFourierProjection, + ImageHintTimeEmbedding, + ImageProjection, + ImageTimeEmbedding, + PositionNet, + TextImageProjection, + TextImageTimeEmbedding, + TextTimeEmbedding, + TimestepEmbedding, + Timesteps, +) + + +from diffusers.models.modeling_utils import ModelMixin +from src.unet_block_hacked_tryon import ( + UNetMidBlock2D, + UNetMidBlock2DCrossAttn, + UNetMidBlock2DSimpleCrossAttn, + get_down_block, + get_up_block, +) +from diffusers.models.resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D +from diffusers.models.transformer_2d import Transformer2DModel +import math + +from ip_adapter.ip_adapter import Resampler + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# def FeedForward(dim, mult=4): +# inner_dim = int(dim * mult) +# return nn.Sequential( +# nn.LayerNorm(dim), +# nn.Linear(dim, inner_dim, bias=False), +# nn.GELU(), +# nn.Linear(inner_dim, dim, bias=False), +# ) + + + +# def reshape_tensor(x, heads): +# bs, length, width = x.shape +# # (bs, length, width) --> (bs, length, n_heads, dim_per_head) +# x = x.view(bs, length, heads, -1) +# # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head) +# x = x.transpose(1, 2) +# # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head) +# x = x.reshape(bs, heads, length, -1) +# return x + + +# class PerceiverAttention(nn.Module): +# def __init__(self, *, dim, dim_head=64, heads=8): +# super().__init__() +# self.scale = dim_head**-0.5 +# self.dim_head = dim_head +# self.heads = heads +# inner_dim = dim_head * heads + +# self.norm1 = nn.LayerNorm(dim) +# self.norm2 = nn.LayerNorm(dim) + +# self.to_q = nn.Linear(dim, inner_dim, bias=False) +# self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False) +# self.to_out = nn.Linear(inner_dim, dim, bias=False) + +# def forward(self, x, latents): +# """ +# Args: +# x (torch.Tensor): image features +# shape (b, n1, D) +# latent (torch.Tensor): latent features +# shape (b, n2, D) +# """ +# x = self.norm1(x) +# latents = self.norm2(latents) + +# b, l, _ = latents.shape + +# q = self.to_q(latents) +# kv_input = torch.cat((x, latents), dim=-2) +# k, v = self.to_kv(kv_input).chunk(2, dim=-1) + +# q = reshape_tensor(q, self.heads) +# k = reshape_tensor(k, self.heads) +# v = reshape_tensor(v, self.heads) + +# # attention +# scale = 1 / math.sqrt(math.sqrt(self.dim_head)) +# weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards +# weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) +# out = weight @ v + +# out = out.permute(0, 2, 1, 3).reshape(b, l, -1) + +# return self.to_out(out) + + +# class Resampler(nn.Module): +# def __init__( +# self, +# dim=1024, +# depth=8, +# dim_head=64, +# heads=16, +# num_queries=8, +# embedding_dim=768, +# output_dim=1024, +# ff_mult=4, +# max_seq_len: int = 257, # CLIP tokens + CLS token +# apply_pos_emb: bool = False, +# num_latents_mean_pooled: int = 0, # number of latents derived from mean pooled representation of the sequence +# ): +# super().__init__() + +# self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5) + +# self.proj_in = nn.Linear(embedding_dim, dim) + +# self.proj_out = nn.Linear(dim, output_dim) +# self.norm_out = nn.LayerNorm(output_dim) + +# self.layers = nn.ModuleList([]) +# for _ in range(depth): +# self.layers.append( +# nn.ModuleList( +# [ +# PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads), +# FeedForward(dim=dim, mult=ff_mult), +# ] +# ) +# ) + +# def forward(self, x): + +# latents = self.latents.repeat(x.size(0), 1, 1) + +# x = self.proj_in(x) + + +# for attn, ff in self.layers: +# latents = attn(x, latents) + latents +# latents = ff(latents) + latents + +# latents = self.proj_out(latents) +# return self.norm_out(latents) + + +def zero_module(module): + for p in module.parameters(): + nn.init.zeros_(p) + return module + +@dataclass +class UNet2DConditionOutput(BaseOutput): + """ + The output of [`UNet2DConditionModel`]. + + Args: + sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): + The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model. + """ + + sample: torch.FloatTensor = None + + +class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin): + r""" + A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample + shaped output. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`): + Height and width of input/output sample. + in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample. + out_channels (`int`, *optional*, defaults to 4): Number of channels in the output. + center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample. + flip_sin_to_cos (`bool`, *optional*, defaults to `False`): + Whether to flip the sin to cos in the time embedding. + freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding. + down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`): + The tuple of downsample blocks to use. + mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`): + Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or + `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`): + The tuple of upsample blocks to use. + only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`): + Whether to include self-attention in the basic transformer blocks, see + [`~models.attention.BasicTransformerBlock`]. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`): + The tuple of output channels for each block. + layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block. + downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution. + mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization. + If `None`, normalization and activation layers is skipped in post-processing. + norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization. + cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280): + The dimension of the cross attention features. + transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling + blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + encoder_hid_dim (`int`, *optional*, defaults to None): + If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim` + dimension to `cross_attention_dim`. + encoder_hid_dim_type (`str`, *optional*, defaults to `None`): + If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text + embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`. + attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads. + num_attention_heads (`int`, *optional*): + The number of attention heads. If not defined, defaults to `attention_head_dim` + resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config + for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`. + class_embed_type (`str`, *optional*, defaults to `None`): + The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`, + `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`. + addition_embed_type (`str`, *optional*, defaults to `None`): + Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or + "text". "text" will use the `TextTimeEmbedding` layer. + addition_time_embed_dim: (`int`, *optional*, defaults to `None`): + Dimension for the timestep embeddings. + num_class_embeds (`int`, *optional*, defaults to `None`): + Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing + class conditioning with `class_embed_type` equal to `None`. + time_embedding_type (`str`, *optional*, defaults to `positional`): + The type of position embedding to use for timesteps. Choose from `positional` or `fourier`. + time_embedding_dim (`int`, *optional*, defaults to `None`): + An optional override for the dimension of the projected time embedding. + time_embedding_act_fn (`str`, *optional*, defaults to `None`): + Optional activation function to use only once on the time embeddings before they are passed to the rest of + the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`. + timestep_post_act (`str`, *optional*, defaults to `None`): + The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`. + time_cond_proj_dim (`int`, *optional*, defaults to `None`): + The dimension of `cond_proj` layer in the timestep embedding. + conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`, + *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`, + *optional*): The dimension of the `class_labels` input when + `class_embed_type="projection"`. Required when `class_embed_type="projection"`. + class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time + embeddings with the class embeddings. + mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`): + Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If + `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the + `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False` + otherwise. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + sample_size: Optional[int] = None, + in_channels: int = 4, + out_channels: int = 4, + center_input_sample: bool = False, + flip_sin_to_cos: bool = True, + freq_shift: int = 0, + down_block_types: Tuple[str] = ( + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "DownBlock2D", + ), + mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn", + up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"), + only_cross_attention: Union[bool, Tuple[bool]] = False, + block_out_channels: Tuple[int] = (320, 640, 1280, 1280), + layers_per_block: Union[int, Tuple[int]] = 2, + downsample_padding: int = 1, + mid_block_scale_factor: float = 1, + dropout: float = 0.0, + act_fn: str = "silu", + norm_num_groups: Optional[int] = 32, + norm_eps: float = 1e-5, + cross_attention_dim: Union[int, Tuple[int]] = 1280, + transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1, + reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None, + encoder_hid_dim: Optional[int] = None, + encoder_hid_dim_type: Optional[str] = None, + attention_head_dim: Union[int, Tuple[int]] = 8, + num_attention_heads: Optional[Union[int, Tuple[int]]] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + class_embed_type: Optional[str] = None, + addition_embed_type: Optional[str] = None, + addition_time_embed_dim: Optional[int] = None, + num_class_embeds: Optional[int] = None, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: int = 1.0, + time_embedding_type: str = "positional", + time_embedding_dim: Optional[int] = None, + time_embedding_act_fn: Optional[str] = None, + timestep_post_act: Optional[str] = None, + time_cond_proj_dim: Optional[int] = None, + conv_in_kernel: int = 3, + conv_out_kernel: int = 3, + projection_class_embeddings_input_dim: Optional[int] = None, + attention_type: str = "default", + class_embeddings_concat: bool = False, + mid_block_only_cross_attention: Optional[bool] = None, + cross_attention_norm: Optional[str] = None, + addition_embed_type_num_heads=64, + ): + super().__init__() + + self.sample_size = sample_size + + if num_attention_heads is not None: + raise ValueError( + "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19." + ) + + # If `num_attention_heads` is not defined (which is the case for most models) + # it will default to `attention_head_dim`. This looks weird upon first reading it and it is. + # The reason for this behavior is to correct for incorrectly named variables that were introduced + # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 + # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking + # which is why we correct for the naming here. + num_attention_heads = num_attention_heads or attention_head_dim + + # Check inputs + if len(down_block_types) != len(up_block_types): + raise ValueError( + f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}." + ) + + if len(block_out_channels) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}." + ) + + if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}." + ) + if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None: + for layer_number_per_block in transformer_layers_per_block: + if isinstance(layer_number_per_block, list): + raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.") + + # input + conv_in_padding = (conv_in_kernel - 1) // 2 + self.conv_in = nn.Conv2d( + in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding + ) + + # time + if time_embedding_type == "fourier": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 2 + if time_embed_dim % 2 != 0: + raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.") + self.time_proj = GaussianFourierProjection( + time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos + ) + timestep_input_dim = time_embed_dim + elif time_embedding_type == "positional": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 4 + + self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift) + timestep_input_dim = block_out_channels[0] + else: + raise ValueError( + f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`." + ) + + self.time_embedding = TimestepEmbedding( + timestep_input_dim, + time_embed_dim, + act_fn=act_fn, + post_act_fn=timestep_post_act, + cond_proj_dim=time_cond_proj_dim, + ) + + if encoder_hid_dim_type is None and encoder_hid_dim is not None: + encoder_hid_dim_type = "text_proj" + self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type) + logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.") + + if encoder_hid_dim is None and encoder_hid_dim_type is not None: + raise ValueError( + f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}." + ) + + if encoder_hid_dim_type == "text_proj": + self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim) + elif encoder_hid_dim_type == "text_image_proj": + # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)` + self.encoder_hid_proj = TextImageProjection( + text_embed_dim=encoder_hid_dim, + image_embed_dim=cross_attention_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 + self.encoder_hid_proj = ImageProjection( + image_embed_dim=encoder_hid_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type == "ip_image_proj": + # Kandinsky 2.2 + self.encoder_hid_proj = Resampler( + dim=1280, + depth=4, + dim_head=64, + heads=20, + num_queries=16, + embedding_dim=encoder_hid_dim, + output_dim=self.config.cross_attention_dim, + ff_mult=4, + ) + + + elif encoder_hid_dim_type is not None: + raise ValueError( + f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'." + ) + else: + self.encoder_hid_proj = None + + # class embedding + if class_embed_type is None and num_class_embeds is not None: + self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim) + elif class_embed_type == "timestep": + self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn) + elif class_embed_type == "identity": + self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim) + elif class_embed_type == "projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set" + ) + # The projection `class_embed_type` is the same as the timestep `class_embed_type` except + # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings + # 2. it projects from an arbitrary input dimension. + # + # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations. + # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings. + # As a result, `TimestepEmbedding` can be passed arbitrary vectors. + self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif class_embed_type == "simple_projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set" + ) + self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim) + else: + self.class_embedding = None + + if addition_embed_type == "text": + if encoder_hid_dim is not None: + text_time_embedding_from_dim = encoder_hid_dim + else: + text_time_embedding_from_dim = cross_attention_dim + + self.add_embedding = TextTimeEmbedding( + text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads + ) + elif addition_embed_type == "text_image": + # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)` + self.add_embedding = TextImageTimeEmbedding( + text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim + ) + elif addition_embed_type == "text_time": + self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift) + self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif addition_embed_type == "image": + # Kandinsky 2.2 + self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type == "image_hint": + # Kandinsky 2.2 ControlNet + self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type is not None: + raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.") + + if time_embedding_act_fn is None: + self.time_embed_act = None + else: + self.time_embed_act = get_activation(time_embedding_act_fn) + + self.down_blocks = nn.ModuleList([]) + self.up_blocks = nn.ModuleList([]) + + if isinstance(only_cross_attention, bool): + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = only_cross_attention + + only_cross_attention = [only_cross_attention] * len(down_block_types) + + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = False + + if isinstance(num_attention_heads, int): + num_attention_heads = (num_attention_heads,) * len(down_block_types) + + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * len(down_block_types) + + if isinstance(cross_attention_dim, int): + cross_attention_dim = (cross_attention_dim,) * len(down_block_types) + + if isinstance(layers_per_block, int): + layers_per_block = [layers_per_block] * len(down_block_types) + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types) + if class_embeddings_concat: + # The time embeddings are concatenated with the class embeddings. The dimension of the + # time embeddings passed to the down, middle, and up blocks is twice the dimension of the + # regular time embeddings + blocks_time_embed_dim = time_embed_dim * 2 + else: + blocks_time_embed_dim = time_embed_dim + + # down + output_channel = block_out_channels[0] + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + + down_block = get_down_block( + down_block_type, + num_layers=layers_per_block[i], + transformer_layers_per_block=transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + temb_channels=blocks_time_embed_dim, + add_downsample=not is_final_block, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim[i], + num_attention_heads=num_attention_heads[i], + downsample_padding=downsample_padding, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.down_blocks.append(down_block) + + # mid + if mid_block_type == "UNetMidBlock2DCrossAttn": + self.mid_block = UNetMidBlock2DCrossAttn( + transformer_layers_per_block=transformer_layers_per_block[-1], + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_time_scale_shift=resnet_time_scale_shift, + cross_attention_dim=cross_attention_dim[-1], + num_attention_heads=num_attention_heads[-1], + resnet_groups=norm_num_groups, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn": + self.mid_block = UNetMidBlock2DSimpleCrossAttn( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + cross_attention_dim=cross_attention_dim[-1], + attention_head_dim=attention_head_dim[-1], + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + only_cross_attention=mid_block_only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif mid_block_type == "UNetMidBlock2D": + self.mid_block = UNetMidBlock2D( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + num_layers=0, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + add_attention=False, + ) + elif mid_block_type is None: + self.mid_block = None + else: + raise ValueError(f"unknown mid_block_type : {mid_block_type}") + + # count how many layers upsample the images + self.num_upsamplers = 0 + + # up + reversed_block_out_channels = list(reversed(block_out_channels)) + reversed_num_attention_heads = list(reversed(num_attention_heads)) + reversed_layers_per_block = list(reversed(layers_per_block)) + reversed_cross_attention_dim = list(reversed(cross_attention_dim)) + reversed_transformer_layers_per_block = ( + list(reversed(transformer_layers_per_block)) + if reverse_transformer_layers_per_block is None + else reverse_transformer_layers_per_block + ) + only_cross_attention = list(reversed(only_cross_attention)) + + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(up_block_types): + is_final_block = i == len(block_out_channels) - 1 + + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)] + + # add upsample block for all BUT final layer + if not is_final_block: + add_upsample = True + self.num_upsamplers += 1 + else: + add_upsample = False + up_block = get_up_block( + up_block_type, + num_layers=reversed_layers_per_block[i] + 1, + transformer_layers_per_block=reversed_transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + prev_output_channel=prev_output_channel, + temb_channels=blocks_time_embed_dim, + add_upsample=add_upsample, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resolution_idx=i, + resnet_groups=norm_num_groups, + cross_attention_dim=reversed_cross_attention_dim[i], + num_attention_heads=reversed_num_attention_heads[i], + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + + + + # out + if norm_num_groups is not None: + self.conv_norm_out = nn.GroupNorm( + num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps + ) + + self.conv_act = get_activation(act_fn) + + else: + self.conv_norm_out = None + self.conv_act = None + + conv_out_padding = (conv_out_kernel - 1) // 2 + self.conv_out = nn.Conv2d( + block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding + ) + + if attention_type in ["gated", "gated-text-image"]: + positive_len = 768 + if isinstance(cross_attention_dim, int): + positive_len = cross_attention_dim + elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list): + positive_len = cross_attention_dim[0] + + feature_type = "text-only" if attention_type == "gated" else "text-image" + self.position_net = PositionNet( + positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type + ) + + + + from ip_adapter.attention_processor import IPAttnProcessor2_0 as IPAttnProcessor, AttnProcessor2_0 as AttnProcessor + + attn_procs = {} + for name in self.attn_processors.keys(): + cross_attention_dim = None if name.endswith("attn1.processor") else self.config.cross_attention_dim + if name.startswith("mid_block"): + hidden_size = self.config.block_out_channels[-1] + elif name.startswith("up_blocks"): + block_id = int(name[len("up_blocks.")]) + hidden_size = list(reversed(self.config.block_out_channels))[block_id] + elif name.startswith("down_blocks"): + block_id = int(name[len("down_blocks.")]) + hidden_size = self.config.block_out_channels[block_id] + if cross_attention_dim is None: + attn_procs[name] = AttnProcessor() + else: + layer_name = name.split(".processor")[0] + attn_procs[name] = IPAttnProcessor(hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, num_tokens=16) + self.set_attn_processor(attn_procs) + + + @property + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True) + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + def set_attn_processor( + self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False + ): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor, _remove_lora=True) + + def set_attention_slice(self, slice_size): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module splits the input tensor in slices to compute attention in + several steps. This is useful for saving some memory in exchange for a small decrease in speed. + + Args: + slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`): + When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If + `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is + provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` + must be a multiple of `slice_size`. + """ + sliceable_head_dims = [] + + def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module): + if hasattr(module, "set_attention_slice"): + sliceable_head_dims.append(module.sliceable_head_dim) + + for child in module.children(): + fn_recursive_retrieve_sliceable_dims(child) + + # retrieve number of attention layers + for module in self.children(): + fn_recursive_retrieve_sliceable_dims(module) + + num_sliceable_layers = len(sliceable_head_dims) + + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = [dim // 2 for dim in sliceable_head_dims] + elif slice_size == "max": + # make smallest slice possible + slice_size = num_sliceable_layers * [1] + + slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size + + if len(slice_size) != len(sliceable_head_dims): + raise ValueError( + f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different" + f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}." + ) + + for i in range(len(slice_size)): + size = slice_size[i] + dim = sliceable_head_dims[i] + if size is not None and size > dim: + raise ValueError(f"size {size} has to be smaller or equal to {dim}.") + + # Recursively walk through all the children. + # Any children which exposes the set_attention_slice method + # gets the message + def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]): + if hasattr(module, "set_attention_slice"): + module.set_attention_slice(slice_size.pop()) + + for child in module.children(): + fn_recursive_set_attention_slice(child, slice_size) + + reversed_slice_size = list(reversed(slice_size)) + for module in self.children(): + fn_recursive_set_attention_slice(module, reversed_slice_size) + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def enable_freeu(self, s1, s2, b1, b2): + r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497. + + The suffixes after the scaling factors represent the stage blocks where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that + are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + for i, upsample_block in enumerate(self.up_blocks): + setattr(upsample_block, "s1", s1) + setattr(upsample_block, "s2", s2) + setattr(upsample_block, "b1", b1) + setattr(upsample_block, "b2", b2) + + def disable_freeu(self): + """Disables the FreeU mechanism.""" + freeu_keys = {"s1", "s2", "b1", "b2"} + for i, upsample_block in enumerate(self.up_blocks): + for k in freeu_keys: + if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None: + setattr(upsample_block, k, None) + + def fuse_qkv_projections(self): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is ๐Ÿงช experimental. + + + """ + self.original_attn_processors = None + + for _, attn_processor in self.attn_processors.items(): + if "Added" in str(attn_processor.__class__.__name__): + raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.") + + self.original_attn_processors = self.attn_processors + + for module in self.modules(): + if isinstance(module, Attention): + module.fuse_projections(fuse=True) + + def unfuse_qkv_projections(self): + """Disables the fused QKV projection if enabled. + + + + This API is ๐Ÿงช experimental. + + + + """ + if self.original_attn_processors is not None: + self.set_attn_processor(self.original_attn_processors) + + def forward( + self, + sample: torch.FloatTensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + class_labels: Optional[torch.Tensor] = None, + timestep_cond: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + mid_block_additional_residual: Optional[torch.Tensor] = None, + down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + garment_features: Optional[Tuple[torch.Tensor]] = None, + ) -> Union[UNet2DConditionOutput, Tuple]: + r""" + The [`UNet2DConditionModel`] forward method. + + Args: + sample (`torch.FloatTensor`): + The noisy input tensor with the following shape `(batch, channel, height, width)`. + timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input. + encoder_hidden_states (`torch.FloatTensor`): + The encoder hidden states with shape `(batch, sequence_length, feature_dim)`. + class_labels (`torch.Tensor`, *optional*, defaults to `None`): + Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings. + timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`): + Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed + through the `self.time_embedding` layer to obtain the timestep embeddings. + attention_mask (`torch.Tensor`, *optional*, defaults to `None`): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*): + A tuple of tensors that if specified are added to the residuals of down unet blocks. + mid_block_additional_residual: (`torch.Tensor`, *optional*): + A tensor that if specified is added to the residual of the middle unet block. + encoder_attention_mask (`torch.Tensor`): + A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If + `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias, + which adds large negative values to the attention scores corresponding to "discard" tokens. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttnProcessor`]. + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added to UNet long skip connections from down blocks to up blocks for + example from ControlNet side model(s) + mid_block_additional_residual (`torch.Tensor`, *optional*): + additional residual to be added to UNet mid block output, for example from ControlNet side model + down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s) + + Returns: + [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`: + If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise + a `tuple` is returned where the first element is the sample tensor. + """ + # By default samples have to be AT least a multiple of the overall upsampling factor. + # The overall upsampling factor is equal to 2 ** (# num of upsampling layers). + # However, the upsampling interpolation output size can be forced to fit any upsampling size + # on the fly if necessary. + default_overall_up_factor = 2**self.num_upsamplers + + # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` + forward_upsample_size = False + upsample_size = None + + for dim in sample.shape[-2:]: + if dim % default_overall_up_factor != 0: + # Forward upsample size to force interpolation output size. + forward_upsample_size = True + break + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None: + encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # 0. center input if necessary + if self.config.center_input_sample: + sample = 2 * sample - 1.0 + + # 1. time + timesteps = timestep + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = sample.device.type == "mps" + if isinstance(timestep, float): + dtype = torch.float32 if is_mps else torch.float64 + else: + dtype = torch.int32 if is_mps else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + t_emb = self.time_proj(timesteps) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=sample.dtype) + + emb = self.time_embedding(t_emb, timestep_cond) + aug_emb = None + + if self.class_embedding is not None: + if class_labels is None: + raise ValueError("class_labels should be provided when num_class_embeds > 0") + + if self.config.class_embed_type == "timestep": + class_labels = self.time_proj(class_labels) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # there might be better ways to encapsulate this. + class_labels = class_labels.to(dtype=sample.dtype) + + class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype) + + if self.config.class_embeddings_concat: + emb = torch.cat([emb, class_emb], dim=-1) + else: + emb = emb + class_emb + + if self.config.addition_embed_type == "text": + aug_emb = self.add_embedding(encoder_hidden_states) + elif self.config.addition_embed_type == "text_image": + # Kandinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + + image_embs = added_cond_kwargs.get("image_embeds") + text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states) + aug_emb = self.add_embedding(text_embs, image_embs) + elif self.config.addition_embed_type == "text_time": + # SDXL - style + if "text_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`" + ) + text_embeds = added_cond_kwargs.get("text_embeds") + if "time_ids" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`" + ) + time_ids = added_cond_kwargs.get("time_ids") + time_embeds = self.add_time_proj(time_ids.flatten()) + time_embeds = time_embeds.reshape((text_embeds.shape[0], -1)) + add_embeds = torch.concat([text_embeds, time_embeds], dim=-1) + add_embeds = add_embeds.to(emb.dtype) + aug_emb = self.add_embedding(add_embeds) + elif self.config.addition_embed_type == "image": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + aug_emb = self.add_embedding(image_embs) + elif self.config.addition_embed_type == "image_hint": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + hint = added_cond_kwargs.get("hint") + aug_emb, hint = self.add_embedding(image_embs, hint) + sample = torch.cat([sample, hint], dim=1) + + emb = emb + aug_emb if aug_emb is not None else emb + + if self.time_embed_act is not None: + emb = self.time_embed_act(emb) + + if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj": + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj": + # Kadinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj": + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + # print(image_embeds.shape) + # image_embeds = self.encoder_hid_proj(image_embeds).to(encoder_hidden_states.dtype) + encoder_hidden_states = torch.cat([encoder_hidden_states, image_embeds], dim=1) + + # 2. pre-process + sample = self.conv_in(sample) + + # 2.5 GLIGEN position net + if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None: + cross_attention_kwargs = cross_attention_kwargs.copy() + gligen_args = cross_attention_kwargs.pop("gligen") + cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)} + + + curr_garment_feat_idx = 0 + + + # 3. down + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + if USE_PEFT_BACKEND: + # weight the lora layers by setting `lora_scale` for each PEFT layer + scale_lora_layers(self, lora_scale) + + is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None + # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets + is_adapter = down_intrablock_additional_residuals is not None + # maintain backward compatibility for legacy usage, where + # T2I-Adapter and ControlNet both use down_block_additional_residuals arg + # but can only use one or the other + if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None: + deprecate( + "T2I should not use down_block_additional_residuals", + "1.3.0", + "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \ + and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \ + for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ", + standard_warn=False, + ) + down_intrablock_additional_residuals = down_block_additional_residuals + is_adapter = True + + down_block_res_samples = (sample,) + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + # For t2i-adapter CrossAttnDownBlock2D + additional_residuals = {} + if is_adapter and len(down_intrablock_additional_residuals) > 0: + additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0) + + sample, res_samples,curr_garment_feat_idx = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + garment_features=garment_features, + curr_garment_feat_idx=curr_garment_feat_idx, + **additional_residuals, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb, scale=lora_scale) + if is_adapter and len(down_intrablock_additional_residuals) > 0: + sample += down_intrablock_additional_residuals.pop(0) + + down_block_res_samples += res_samples + + + if is_controlnet: + new_down_block_res_samples = () + + for down_block_res_sample, down_block_additional_residual in zip( + down_block_res_samples, down_block_additional_residuals + ): + down_block_res_sample = down_block_res_sample + down_block_additional_residual + new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = new_down_block_res_samples + + # 4. mid + if self.mid_block is not None: + if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention: + sample ,curr_garment_feat_idx= self.mid_block( + sample, + emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + garment_features=garment_features, + curr_garment_feat_idx=curr_garment_feat_idx, + ) + else: + sample = self.mid_block(sample, emb) + + # To support T2I-Adapter-XL + if ( + is_adapter + and len(down_intrablock_additional_residuals) > 0 + and sample.shape == down_intrablock_additional_residuals[0].shape + ): + sample += down_intrablock_additional_residuals.pop(0) + + if is_controlnet: + sample = sample + mid_block_additional_residual + + + + # 5. up + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] + + # if we have not reached the final block and need to forward the + # upsample size, we do it here + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: + sample ,curr_garment_feat_idx= upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + upsample_size=upsample_size, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + garment_features=garment_features, + curr_garment_feat_idx=curr_garment_feat_idx, + ) + + else: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + upsample_size=upsample_size, + scale=lora_scale, + ) + # 6. post-process + if self.conv_norm_out: + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + if USE_PEFT_BACKEND: + # remove `lora_scale` from each PEFT layer + unscale_lora_layers(self, lora_scale) + + if not return_dict: + return (sample,) + + return UNet2DConditionOutput(sample=sample)