diff --git a/extensions-builtin/sd_forge_ipadapter/IPAdapterPlus.py b/extensions-builtin/sd_forge_ipadapter/IPAdapterPlus.py new file mode 100644 index 00000000..435d9241 --- /dev/null +++ b/extensions-builtin/sd_forge_ipadapter/IPAdapterPlus.py @@ -0,0 +1,943 @@ +import torch +import contextlib +import os +import math + +import ldm_patched.modules.utils +import ldm_patched.modules.model_management +from ldm_patched.modules.clip_vision import clip_preprocess +from ldm_patched.ldm.modules.attention import optimized_attention +from modules_forge.shared import preprocessor_dir + +from torch import nn +from PIL import Image +import torch.nn.functional as F +import torchvision.transforms as TT + +from .resampler import PerceiverAttention, FeedForward, Resampler + + +GLOBAL_MODELS_DIR = preprocessor_dir +MODELS_DIR = preprocessor_dir +INSIGHTFACE_DIR = os.path.join(preprocessor_dir, "insightface") + +os.makedirs(INSIGHTFACE_DIR, exist_ok=True) + + +class FacePerceiverResampler(torch.nn.Module): + def __init__( + self, + *, + dim=768, + depth=4, + dim_head=64, + heads=16, + embedding_dim=1280, + output_dim=768, + ff_mult=4, + ): + super().__init__() + + self.proj_in = torch.nn.Linear(embedding_dim, dim) + self.proj_out = torch.nn.Linear(dim, output_dim) + self.norm_out = torch.nn.LayerNorm(output_dim) + self.layers = torch.nn.ModuleList([]) + for _ in range(depth): + self.layers.append( + torch.nn.ModuleList( + [ + PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads), + FeedForward(dim=dim, mult=ff_mult), + ] + ) + ) + + def forward(self, latents, x): + 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) + +class MLPProjModel(torch.nn.Module): + def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024): + super().__init__() + + self.proj = torch.nn.Sequential( + torch.nn.Linear(clip_embeddings_dim, clip_embeddings_dim), + torch.nn.GELU(), + torch.nn.Linear(clip_embeddings_dim, cross_attention_dim), + torch.nn.LayerNorm(cross_attention_dim) + ) + + def forward(self, image_embeds): + clip_extra_context_tokens = self.proj(image_embeds) + return clip_extra_context_tokens + +class MLPProjModelFaceId(torch.nn.Module): + def __init__(self, cross_attention_dim=768, id_embeddings_dim=512, num_tokens=4): + super().__init__() + + self.cross_attention_dim = cross_attention_dim + self.num_tokens = num_tokens + + self.proj = torch.nn.Sequential( + torch.nn.Linear(id_embeddings_dim, id_embeddings_dim*2), + torch.nn.GELU(), + torch.nn.Linear(id_embeddings_dim*2, cross_attention_dim*num_tokens), + ) + self.norm = torch.nn.LayerNorm(cross_attention_dim) + + def forward(self, id_embeds): + clip_extra_context_tokens = self.proj(id_embeds) + clip_extra_context_tokens = clip_extra_context_tokens.reshape(-1, self.num_tokens, self.cross_attention_dim) + clip_extra_context_tokens = self.norm(clip_extra_context_tokens) + return clip_extra_context_tokens + +class ProjModelFaceIdPlus(torch.nn.Module): + def __init__(self, cross_attention_dim=768, id_embeddings_dim=512, clip_embeddings_dim=1280, num_tokens=4): + super().__init__() + + self.cross_attention_dim = cross_attention_dim + self.num_tokens = num_tokens + + self.proj = torch.nn.Sequential( + torch.nn.Linear(id_embeddings_dim, id_embeddings_dim*2), + torch.nn.GELU(), + torch.nn.Linear(id_embeddings_dim*2, cross_attention_dim*num_tokens), + ) + self.norm = torch.nn.LayerNorm(cross_attention_dim) + + self.perceiver_resampler = FacePerceiverResampler( + dim=cross_attention_dim, + depth=4, + dim_head=64, + heads=cross_attention_dim // 64, + embedding_dim=clip_embeddings_dim, + output_dim=cross_attention_dim, + ff_mult=4, + ) + + def forward(self, id_embeds, clip_embeds, scale=1.0, shortcut=False): + x = self.proj(id_embeds) + x = x.reshape(-1, self.num_tokens, self.cross_attention_dim) + x = self.norm(x) + out = self.perceiver_resampler(x, clip_embeds) + if shortcut: + out = x + scale * out + return out + +class ImageProjModel(nn.Module): + def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4): + super().__init__() + + self.cross_attention_dim = cross_attention_dim + self.clip_extra_context_tokens = clip_extra_context_tokens + self.proj = nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim) + self.norm = nn.LayerNorm(cross_attention_dim) + + def forward(self, image_embeds): + embeds = image_embeds + clip_extra_context_tokens = self.proj(embeds).reshape(-1, self.clip_extra_context_tokens, self.cross_attention_dim) + clip_extra_context_tokens = self.norm(clip_extra_context_tokens) + return clip_extra_context_tokens + +class To_KV(nn.Module): + def __init__(self, state_dict): + super().__init__() + + self.to_kvs = nn.ModuleDict() + for key, value in state_dict.items(): + self.to_kvs[key.replace(".weight", "").replace(".", "_")] = nn.Linear(value.shape[1], value.shape[0], bias=False) + self.to_kvs[key.replace(".weight", "").replace(".", "_")].weight.data = value + +def set_model_patch_replace(model, patch_kwargs, key): + to = model.model_options["transformer_options"] + if "patches_replace" not in to: + to["patches_replace"] = {} + if "attn2" not in to["patches_replace"]: + to["patches_replace"]["attn2"] = {} + if key not in to["patches_replace"]["attn2"]: + patch = CrossAttentionPatch(**patch_kwargs) + to["patches_replace"]["attn2"][key] = patch + else: + to["patches_replace"]["attn2"][key].set_new_condition(**patch_kwargs) + +def image_add_noise(image, noise): + image = image.permute([0,3,1,2]) + torch.manual_seed(0) # use a fixed random for reproducible results + transforms = TT.Compose([ + TT.CenterCrop(min(image.shape[2], image.shape[3])), + TT.Resize((224, 224), interpolation=TT.InterpolationMode.BICUBIC, antialias=True), + TT.ElasticTransform(alpha=75.0, sigma=noise*3.5), # shuffle the image + TT.RandomVerticalFlip(p=1.0), # flip the image to change the geometry even more + TT.RandomHorizontalFlip(p=1.0), + ]) + image = transforms(image.cpu()) + image = image.permute([0,2,3,1]) + image = image + ((0.25*(1-noise)+0.05) * torch.randn_like(image) ) # add further random noise + return image + +def zeroed_hidden_states(clip_vision, batch_size): + image = torch.zeros([batch_size, 224, 224, 3]) + ldm_patched.modules.model_management.load_model_gpu(clip_vision.patcher) + pixel_values = clip_preprocess(image.to(clip_vision.load_device)).float() + outputs = clip_vision.model(pixel_values=pixel_values, intermediate_output=-2) + # we only need the penultimate hidden states + outputs = outputs[1].to(ldm_patched.modules.model_management.intermediate_device()) + return outputs + +def min_(tensor_list): + # return the element-wise min of the tensor list. + x = torch.stack(tensor_list) + mn = x.min(axis=0)[0] + return torch.clamp(mn, min=0) + +def max_(tensor_list): + # return the element-wise max of the tensor list. + x = torch.stack(tensor_list) + mx = x.max(axis=0)[0] + return torch.clamp(mx, max=1) + +# From https://github.com/Jamy-L/Pytorch-Contrast-Adaptive-Sharpening/ +def contrast_adaptive_sharpening(image, amount): + img = F.pad(image, pad=(1, 1, 1, 1)).cpu() + + a = img[..., :-2, :-2] + b = img[..., :-2, 1:-1] + c = img[..., :-2, 2:] + d = img[..., 1:-1, :-2] + e = img[..., 1:-1, 1:-1] + f = img[..., 1:-1, 2:] + g = img[..., 2:, :-2] + h = img[..., 2:, 1:-1] + i = img[..., 2:, 2:] + + # Computing contrast + cross = (b, d, e, f, h) + mn = min_(cross) + mx = max_(cross) + + diag = (a, c, g, i) + mn2 = min_(diag) + mx2 = max_(diag) + mx = mx + mx2 + mn = mn + mn2 + + # Computing local weight + inv_mx = torch.reciprocal(mx) + amp = inv_mx * torch.minimum(mn, (2 - mx)) + + # scaling + amp = torch.sqrt(amp) + w = - amp * (amount * (1/5 - 1/8) + 1/8) + div = torch.reciprocal(1 + 4*w) + + output = ((b + d + f + h)*w + e) * div + output = output.clamp(0, 1) + output = torch.nan_to_num(output) + + return (output) + +def tensorToNP(image): + out = torch.clamp(255. * image.detach().cpu(), 0, 255).to(torch.uint8) + out = out[..., [2, 1, 0]] + out = out.numpy() + + return out + +def NPToTensor(image): + out = torch.from_numpy(image) + out = torch.clamp(out.to(torch.float)/255., 0.0, 1.0) + out = out[..., [2, 1, 0]] + + return out + +class IPAdapter(nn.Module): + def __init__(self, ipadapter_model, cross_attention_dim=1024, output_cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4, is_sdxl=False, is_plus=False, is_full=False, is_faceid=False): + super().__init__() + + self.clip_embeddings_dim = clip_embeddings_dim + self.cross_attention_dim = cross_attention_dim + self.output_cross_attention_dim = output_cross_attention_dim + self.clip_extra_context_tokens = clip_extra_context_tokens + self.is_sdxl = is_sdxl + self.is_full = is_full + self.is_plus = is_plus + + if is_faceid: + self.image_proj_model = self.init_proj_faceid() + elif is_plus: + self.image_proj_model = self.init_proj_plus() + else: + self.image_proj_model = self.init_proj() + + self.image_proj_model.load_state_dict(ipadapter_model["image_proj"]) + self.ip_layers = To_KV(ipadapter_model["ip_adapter"]) + + def init_proj(self): + image_proj_model = ImageProjModel( + cross_attention_dim=self.cross_attention_dim, + clip_embeddings_dim=self.clip_embeddings_dim, + clip_extra_context_tokens=self.clip_extra_context_tokens + ) + return image_proj_model + + def init_proj_plus(self): + if self.is_full: + image_proj_model = MLPProjModel( + cross_attention_dim=self.cross_attention_dim, + clip_embeddings_dim=self.clip_embeddings_dim + ) + else: + image_proj_model = Resampler( + dim=self.cross_attention_dim, + depth=4, + dim_head=64, + heads=20 if self.is_sdxl else 12, + num_queries=self.clip_extra_context_tokens, + embedding_dim=self.clip_embeddings_dim, + output_dim=self.output_cross_attention_dim, + ff_mult=4 + ) + return image_proj_model + + def init_proj_faceid(self): + if self.is_plus: + image_proj_model = ProjModelFaceIdPlus( + cross_attention_dim=self.cross_attention_dim, + id_embeddings_dim=512, + clip_embeddings_dim=1280, + num_tokens=4, + ) + else: + image_proj_model = MLPProjModelFaceId( + cross_attention_dim=self.cross_attention_dim, + id_embeddings_dim=512, + num_tokens=self.clip_extra_context_tokens, + ) + return image_proj_model + + @torch.inference_mode() + def get_image_embeds(self, clip_embed, clip_embed_zeroed): + image_prompt_embeds = self.image_proj_model(clip_embed) + uncond_image_prompt_embeds = self.image_proj_model(clip_embed_zeroed) + return image_prompt_embeds, uncond_image_prompt_embeds + + @torch.inference_mode() + def get_image_embeds_faceid_plus(self, face_embed, clip_embed, s_scale, shortcut): + embeds = self.image_proj_model(face_embed, clip_embed, scale=s_scale, shortcut=shortcut) + return embeds + +class CrossAttentionPatch: + # forward for patching + def __init__(self, weight, ipadapter, number, cond, uncond, weight_type, mask=None, sigma_start=0.0, sigma_end=1.0, unfold_batch=False): + self.weights = [weight] + self.ipadapters = [ipadapter] + self.conds = [cond] + self.unconds = [uncond] + self.number = number + self.weight_type = [weight_type] + self.masks = [mask] + self.sigma_start = [sigma_start] + self.sigma_end = [sigma_end] + self.unfold_batch = [unfold_batch] + + self.k_key = str(self.number*2+1) + "_to_k_ip" + self.v_key = str(self.number*2+1) + "_to_v_ip" + + def set_new_condition(self, weight, ipadapter, number, cond, uncond, weight_type, mask=None, sigma_start=0.0, sigma_end=1.0, unfold_batch=False): + self.weights.append(weight) + self.ipadapters.append(ipadapter) + self.conds.append(cond) + self.unconds.append(uncond) + self.masks.append(mask) + self.weight_type.append(weight_type) + self.sigma_start.append(sigma_start) + self.sigma_end.append(sigma_end) + self.unfold_batch.append(unfold_batch) + + def __call__(self, n, context_attn2, value_attn2, extra_options): + org_dtype = n.dtype + cond_or_uncond = extra_options["cond_or_uncond"] + sigma = extra_options["sigmas"][0].item() if 'sigmas' in extra_options else 999999999.9 + + # extra options for AnimateDiff + ad_params = extra_options['ad_params'] if "ad_params" in extra_options else None + + q = n + k = context_attn2 + v = value_attn2 + b = q.shape[0] + qs = q.shape[1] + batch_prompt = b // len(cond_or_uncond) + out = optimized_attention(q, k, v, extra_options["n_heads"]) + _, _, lh, lw = extra_options["original_shape"] + + for weight, cond, uncond, ipadapter, mask, weight_type, sigma_start, sigma_end, unfold_batch in zip(self.weights, self.conds, self.unconds, self.ipadapters, self.masks, self.weight_type, self.sigma_start, self.sigma_end, self.unfold_batch): + if sigma > sigma_start or sigma < sigma_end: + continue + + if unfold_batch and cond.shape[0] > 1: + # Check AnimateDiff context window + if ad_params is not None and ad_params["sub_idxs"] is not None: + # if images length matches or exceeds full_length get sub_idx images + if cond.shape[0] >= ad_params["full_length"]: + cond = torch.Tensor(cond[ad_params["sub_idxs"]]) + uncond = torch.Tensor(uncond[ad_params["sub_idxs"]]) + # otherwise, need to do more to get proper sub_idxs masks + else: + # check if images length matches full_length - if not, make it match + if cond.shape[0] < ad_params["full_length"]: + cond = torch.cat((cond, cond[-1:].repeat((ad_params["full_length"]-cond.shape[0], 1, 1))), dim=0) + uncond = torch.cat((uncond, uncond[-1:].repeat((ad_params["full_length"]-uncond.shape[0], 1, 1))), dim=0) + # if we have too many remove the excess (should not happen, but just in case) + if cond.shape[0] > ad_params["full_length"]: + cond = cond[:ad_params["full_length"]] + uncond = uncond[:ad_params["full_length"]] + cond = cond[ad_params["sub_idxs"]] + uncond = uncond[ad_params["sub_idxs"]] + + # if we don't have enough reference images repeat the last one until we reach the right size + if cond.shape[0] < batch_prompt: + cond = torch.cat((cond, cond[-1:].repeat((batch_prompt-cond.shape[0], 1, 1))), dim=0) + uncond = torch.cat((uncond, uncond[-1:].repeat((batch_prompt-uncond.shape[0], 1, 1))), dim=0) + # if we have too many remove the exceeding + elif cond.shape[0] > batch_prompt: + cond = cond[:batch_prompt] + uncond = uncond[:batch_prompt] + + k_cond = ipadapter.ip_layers.to_kvs[self.k_key](cond) + k_uncond = ipadapter.ip_layers.to_kvs[self.k_key](uncond) + v_cond = ipadapter.ip_layers.to_kvs[self.v_key](cond) + v_uncond = ipadapter.ip_layers.to_kvs[self.v_key](uncond) + else: + k_cond = ipadapter.ip_layers.to_kvs[self.k_key](cond).repeat(batch_prompt, 1, 1) + k_uncond = ipadapter.ip_layers.to_kvs[self.k_key](uncond).repeat(batch_prompt, 1, 1) + v_cond = ipadapter.ip_layers.to_kvs[self.v_key](cond).repeat(batch_prompt, 1, 1) + v_uncond = ipadapter.ip_layers.to_kvs[self.v_key](uncond).repeat(batch_prompt, 1, 1) + + if weight_type.startswith("linear"): + ip_k = torch.cat([(k_cond, k_uncond)[i] for i in cond_or_uncond], dim=0) * weight + ip_v = torch.cat([(v_cond, v_uncond)[i] for i in cond_or_uncond], dim=0) * weight + else: + ip_k = torch.cat([(k_cond, k_uncond)[i] for i in cond_or_uncond], dim=0) + ip_v = torch.cat([(v_cond, v_uncond)[i] for i in cond_or_uncond], dim=0) + + if weight_type.startswith("channel"): + # code by Lvmin Zhang at Stanford University as also seen on Fooocus IPAdapter implementation + # please read licensing notes https://github.com/lllyasviel/Fooocus/blob/69a23c4d60c9e627409d0cb0f8862cdb015488eb/extras/ip_adapter.py#L234 + ip_v_mean = torch.mean(ip_v, dim=1, keepdim=True) + ip_v_offset = ip_v - ip_v_mean + _, _, C = ip_k.shape + channel_penalty = float(C) / 1280.0 + W = weight * channel_penalty + ip_k = ip_k * W + ip_v = ip_v_offset + ip_v_mean * W + + out_ip = optimized_attention(q, ip_k, ip_v, extra_options["n_heads"]) + if weight_type.startswith("original"): + out_ip = out_ip * weight + + if mask is not None: + # TODO: needs checking + mask_h = lh / math.sqrt(lh * lw / qs) + mask_h = int(mask_h) + int((qs % int(mask_h)) != 0) + mask_w = qs // mask_h + + # check if using AnimateDiff and sliding context window + if (mask.shape[0] > 1 and ad_params is not None and ad_params["sub_idxs"] is not None): + # if mask length matches or exceeds full_length, just get sub_idx masks, resize, and continue + if mask.shape[0] >= ad_params["full_length"]: + mask_downsample = torch.Tensor(mask[ad_params["sub_idxs"]]) + mask_downsample = F.interpolate(mask_downsample.unsqueeze(1), size=(mask_h, mask_w), mode="bicubic").squeeze(1) + # otherwise, need to do more to get proper sub_idxs masks + else: + # resize to needed attention size (to save on memory) + mask_downsample = F.interpolate(mask.unsqueeze(1), size=(mask_h, mask_w), mode="bicubic").squeeze(1) + # check if mask length matches full_length - if not, make it match + if mask_downsample.shape[0] < ad_params["full_length"]: + mask_downsample = torch.cat((mask_downsample, mask_downsample[-1:].repeat((ad_params["full_length"]-mask_downsample.shape[0], 1, 1))), dim=0) + # if we have too many remove the excess (should not happen, but just in case) + if mask_downsample.shape[0] > ad_params["full_length"]: + mask_downsample = mask_downsample[:ad_params["full_length"]] + # now, select sub_idxs masks + mask_downsample = mask_downsample[ad_params["sub_idxs"]] + # otherwise, perform usual mask interpolation + else: + mask_downsample = F.interpolate(mask.unsqueeze(1), size=(mask_h, mask_w), mode="bicubic").squeeze(1) + + # if we don't have enough masks repeat the last one until we reach the right size + if mask_downsample.shape[0] < batch_prompt: + mask_downsample = torch.cat((mask_downsample, mask_downsample[-1:, :, :].repeat((batch_prompt-mask_downsample.shape[0], 1, 1))), dim=0) + # if we have too many remove the exceeding + elif mask_downsample.shape[0] > batch_prompt: + mask_downsample = mask_downsample[:batch_prompt, :, :] + + # repeat the masks + mask_downsample = mask_downsample.repeat(len(cond_or_uncond), 1, 1) + mask_downsample = mask_downsample.view(mask_downsample.shape[0], -1, 1).repeat(1, 1, out.shape[2]) + + out_ip = out_ip * mask_downsample + + out = out + out_ip + + return out.to(dtype=org_dtype) + +class IPAdapterModelLoader: + + RETURN_TYPES = ("IPADAPTER",) + FUNCTION = "load_ipadapter_model" + CATEGORY = "ipadapter" + + def load_ipadapter_model(self, ckpt_path): + model = ldm_patched.modules.utils.load_torch_file(ckpt_path, safe_load=True) + + if ckpt_path.lower().endswith(".safetensors"): + st_model = {"image_proj": {}, "ip_adapter": {}} + for key in model.keys(): + if key.startswith("image_proj."): + st_model["image_proj"][key.replace("image_proj.", "")] = model[key] + elif key.startswith("ip_adapter."): + st_model["ip_adapter"][key.replace("ip_adapter.", "")] = model[key] + model = st_model + + if not "ip_adapter" in model.keys() or not model["ip_adapter"]: + raise Exception("invalid IPAdapter model {}".format(ckpt_path)) + + return (model,) + +insightface_face_align = None +class InsightFaceLoader: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "provider": (["CPU", "CUDA", "ROCM"], ), + }, + } + + RETURN_TYPES = ("INSIGHTFACE",) + FUNCTION = "load_insight_face" + CATEGORY = "ipadapter" + + def load_insight_face(self, provider): + try: + from insightface.app import FaceAnalysis + except ImportError as e: + raise Exception(e) + + from insightface.utils import face_align + global insightface_face_align + insightface_face_align = face_align + + model = FaceAnalysis(name="buffalo_l", root=INSIGHTFACE_DIR, providers=[provider + 'ExecutionProvider',]) + model.prepare(ctx_id=0, det_size=(640, 640)) + + return (model,) + +class IPAdapterApply: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "ipadapter": ("IPADAPTER", ), + "clip_vision": ("CLIP_VISION",), + "image": ("IMAGE",), + "model": ("MODEL", ), + "weight": ("FLOAT", { "default": 1.0, "min": -1, "max": 3, "step": 0.05 }), + "noise": ("FLOAT", { "default": 0.0, "min": 0.0, "max": 1.0, "step": 0.01 }), + "weight_type": (["original", "linear", "channel penalty"], ), + "start_at": ("FLOAT", { "default": 0.0, "min": 0.0, "max": 1.0, "step": 0.001 }), + "end_at": ("FLOAT", { "default": 1.0, "min": 0.0, "max": 1.0, "step": 0.001 }), + "unfold_batch": ("BOOLEAN", { "default": False }), + }, + "optional": { + "attn_mask": ("MASK",), + } + } + + RETURN_TYPES = ("MODEL", ) + FUNCTION = "apply_ipadapter" + CATEGORY = "ipadapter" + + def apply_ipadapter(self, ipadapter, model, weight, clip_vision=None, image=None, weight_type="original", noise=None, embeds=None, attn_mask=None, start_at=0.0, end_at=1.0, unfold_batch=False, insightface=None, faceid_v2=False, weight_v2=False): + self.dtype = torch.float16 if ldm_patched.modules.model_management.should_use_fp16() else torch.float32 + self.device = ldm_patched.modules.model_management.get_torch_device() + self.weight = weight + self.is_full = "proj.3.weight" in ipadapter["image_proj"] + self.is_portrait = "proj.2.weight" in ipadapter["image_proj"] and not "proj.3.weight" in ipadapter["image_proj"] and not "0.to_q_lora.down.weight" in ipadapter["ip_adapter"] + self.is_faceid = self.is_portrait or "0.to_q_lora.down.weight" in ipadapter["ip_adapter"] + self.is_plus = (self.is_full or "latents" in ipadapter["image_proj"] or "perceiver_resampler.proj_in.weight" in ipadapter["image_proj"]) + + if self.is_faceid and not insightface: + raise Exception('InsightFace must be provided for FaceID models.') + + output_cross_attention_dim = ipadapter["ip_adapter"]["1.to_k_ip.weight"].shape[1] + self.is_sdxl = output_cross_attention_dim == 2048 + cross_attention_dim = 1280 if self.is_plus and self.is_sdxl and not self.is_faceid else output_cross_attention_dim + clip_extra_context_tokens = 16 if self.is_plus or self.is_portrait else 4 + + if embeds is not None: + embeds = torch.unbind(embeds) + clip_embed = embeds[0].cpu() + clip_embed_zeroed = embeds[1].cpu() + else: + if self.is_faceid: + insightface.det_model.input_size = (640,640) # reset the detection size + face_img = tensorToNP(image) + face_embed = [] + face_clipvision = [] + + for i in range(face_img.shape[0]): + for size in [(size, size) for size in range(640, 128, -64)]: + insightface.det_model.input_size = size # TODO: hacky but seems to be working + face = insightface.get(face_img[i]) + if face: + face_embed.append(torch.from_numpy(face[0].normed_embedding).unsqueeze(0)) + face_clipvision.append(NPToTensor(insightface_face_align.norm_crop(face_img[i], landmark=face[0].kps, image_size=224))) + + if 640 not in size: + print(f"\033[33mINFO: InsightFace detection resolution lowered to {size}.\033[0m") + break + else: + raise Exception('InsightFace: No face detected.') + + face_embed = torch.stack(face_embed, dim=0) + image = torch.stack(face_clipvision, dim=0) + + neg_image = image_add_noise(image, noise) if noise > 0 else None + + if self.is_plus: + clip_embed = clip_vision.encode_image(image).penultimate_hidden_states + if noise > 0: + clip_embed_zeroed = clip_vision.encode_image(neg_image).penultimate_hidden_states + else: + clip_embed_zeroed = zeroed_hidden_states(clip_vision, image.shape[0]) + + # TODO: check noise to the uncods too + face_embed_zeroed = torch.zeros_like(face_embed) + else: + clip_embed = face_embed + clip_embed_zeroed = torch.zeros_like(clip_embed) + else: + if image.shape[1] != image.shape[2]: + print("\033[33mINFO: the IPAdapter reference image is not a square, CLIPImageProcessor will resize and crop it at the center. If the main focus of the picture is not in the middle the result might not be what you are expecting.\033[0m") + + clip_embed = clip_vision.encode_image(image) + neg_image = image_add_noise(image, noise) if noise > 0 else None + + if self.is_plus: + clip_embed = clip_embed.penultimate_hidden_states + if noise > 0: + clip_embed_zeroed = clip_vision.encode_image(neg_image).penultimate_hidden_states + else: + clip_embed_zeroed = zeroed_hidden_states(clip_vision, image.shape[0]) + else: + clip_embed = clip_embed.image_embeds + if noise > 0: + clip_embed_zeroed = clip_vision.encode_image(neg_image).image_embeds + else: + clip_embed_zeroed = torch.zeros_like(clip_embed) + + clip_embeddings_dim = clip_embed.shape[-1] + + self.ipadapter = IPAdapter( + ipadapter, + cross_attention_dim=cross_attention_dim, + output_cross_attention_dim=output_cross_attention_dim, + clip_embeddings_dim=clip_embeddings_dim, + clip_extra_context_tokens=clip_extra_context_tokens, + is_sdxl=self.is_sdxl, + is_plus=self.is_plus, + is_full=self.is_full, + is_faceid=self.is_faceid, + ) + + self.ipadapter.to(self.device, dtype=self.dtype) + + if self.is_faceid and self.is_plus: + image_prompt_embeds = self.ipadapter.get_image_embeds_faceid_plus(face_embed.to(self.device, dtype=self.dtype), clip_embed.to(self.device, dtype=self.dtype), weight_v2, faceid_v2) + uncond_image_prompt_embeds = self.ipadapter.get_image_embeds_faceid_plus(face_embed_zeroed.to(self.device, dtype=self.dtype), clip_embed_zeroed.to(self.device, dtype=self.dtype), weight_v2, faceid_v2) + else: + image_prompt_embeds, uncond_image_prompt_embeds = self.ipadapter.get_image_embeds(clip_embed.to(self.device, dtype=self.dtype), clip_embed_zeroed.to(self.device, dtype=self.dtype)) + + image_prompt_embeds = image_prompt_embeds.to(self.device, dtype=self.dtype) + uncond_image_prompt_embeds = uncond_image_prompt_embeds.to(self.device, dtype=self.dtype) + + work_model = model.clone() + + if attn_mask is not None: + attn_mask = attn_mask.to(self.device) + + sigma_start = model.model.model_sampling.percent_to_sigma(start_at) + sigma_end = model.model.model_sampling.percent_to_sigma(end_at) + + patch_kwargs = { + "number": 0, + "weight": self.weight, + "ipadapter": self.ipadapter, + "cond": image_prompt_embeds, + "uncond": uncond_image_prompt_embeds, + "weight_type": weight_type, + "mask": attn_mask, + "sigma_start": sigma_start, + "sigma_end": sigma_end, + "unfold_batch": unfold_batch, + } + + if not self.is_sdxl: + for id in [1,2,4,5,7,8]: # id of input_blocks that have cross attention + set_model_patch_replace(work_model, patch_kwargs, ("input", id)) + patch_kwargs["number"] += 1 + for id in [3,4,5,6,7,8,9,10,11]: # id of output_blocks that have cross attention + set_model_patch_replace(work_model, patch_kwargs, ("output", id)) + patch_kwargs["number"] += 1 + set_model_patch_replace(work_model, patch_kwargs, ("middle", 0)) + else: + for id in [4,5,7,8]: # id of input_blocks that have cross attention + block_indices = range(2) if id in [4, 5] else range(10) # transformer_depth + for index in block_indices: + set_model_patch_replace(work_model, patch_kwargs, ("input", id, index)) + patch_kwargs["number"] += 1 + for id in range(6): # id of output_blocks that have cross attention + block_indices = range(2) if id in [3, 4, 5] else range(10) # transformer_depth + for index in block_indices: + set_model_patch_replace(work_model, patch_kwargs, ("output", id, index)) + patch_kwargs["number"] += 1 + for index in range(10): + set_model_patch_replace(work_model, patch_kwargs, ("middle", 0, index)) + patch_kwargs["number"] += 1 + + return (work_model, ) + +class IPAdapterApplyFaceID(IPAdapterApply): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "ipadapter": ("IPADAPTER", ), + "clip_vision": ("CLIP_VISION",), + "insightface": ("INSIGHTFACE",), + "image": ("IMAGE",), + "model": ("MODEL", ), + "weight": ("FLOAT", { "default": 1.0, "min": -1, "max": 3, "step": 0.05 }), + "noise": ("FLOAT", { "default": 0.0, "min": 0.0, "max": 1.0, "step": 0.01 }), + "weight_type": (["original", "linear", "channel penalty"], ), + "start_at": ("FLOAT", { "default": 0.0, "min": 0.0, "max": 1.0, "step": 0.001 }), + "end_at": ("FLOAT", { "default": 1.0, "min": 0.0, "max": 1.0, "step": 0.001 }), + "faceid_v2": ("BOOLEAN", { "default": False }), + "weight_v2": ("FLOAT", { "default": 1.0, "min": -1, "max": 3, "step": 0.05 }), + "unfold_batch": ("BOOLEAN", { "default": False }), + }, + "optional": { + "attn_mask": ("MASK",), + } + } + +def prepImage(image, interpolation="LANCZOS", crop_position="center", size=(224,224), sharpening=0.0, padding=0): + _, oh, ow, _ = image.shape + output = image.permute([0,3,1,2]) + + if "pad" in crop_position: + target_length = max(oh, ow) + pad_l = (target_length - ow) // 2 + pad_r = (target_length - ow) - pad_l + pad_t = (target_length - oh) // 2 + pad_b = (target_length - oh) - pad_t + output = F.pad(output, (pad_l, pad_r, pad_t, pad_b), value=0, mode="constant") + else: + crop_size = min(oh, ow) + x = (ow-crop_size) // 2 + y = (oh-crop_size) // 2 + if "top" in crop_position: + y = 0 + elif "bottom" in crop_position: + y = oh-crop_size + elif "left" in crop_position: + x = 0 + elif "right" in crop_position: + x = ow-crop_size + + x2 = x+crop_size + y2 = y+crop_size + + # crop + output = output[:, :, y:y2, x:x2] + + # resize (apparently PIL resize is better than tourchvision interpolate) + imgs = [] + for i in range(output.shape[0]): + img = TT.ToPILImage()(output[i]) + img = img.resize(size, resample=Image.Resampling[interpolation]) + imgs.append(TT.ToTensor()(img)) + output = torch.stack(imgs, dim=0) + imgs = None # zelous GC + + if sharpening > 0: + output = contrast_adaptive_sharpening(output, sharpening) + + if padding > 0: + output = F.pad(output, (padding, padding, padding, padding), value=255, mode="constant") + + output = output.permute([0,2,3,1]) + + return output + +class PrepImageForInsightFace: + @classmethod + def INPUT_TYPES(s): + return {"required": { + "image": ("IMAGE",), + "crop_position": (["center", "top", "bottom", "left", "right"],), + "sharpening": ("FLOAT", {"default": 0.0, "min": 0, "max": 1, "step": 0.05}), + "pad_around": ("BOOLEAN", { "default": True }), + }, + } + + RETURN_TYPES = ("IMAGE",) + FUNCTION = "prep_image" + + CATEGORY = "ipadapter" + + def prep_image(self, image, crop_position, sharpening=0.0, pad_around=True): + if pad_around: + padding = 30 + size = (580, 580) + else: + padding = 0 + size = (640, 640) + output = prepImage(image, "LANCZOS", crop_position, size, sharpening, padding) + + return (output, ) + +class PrepImageForClipVision: + @classmethod + def INPUT_TYPES(s): + return {"required": { + "image": ("IMAGE",), + "interpolation": (["LANCZOS", "BICUBIC", "HAMMING", "BILINEAR", "BOX", "NEAREST"],), + "crop_position": (["top", "bottom", "left", "right", "center", "pad"],), + "sharpening": ("FLOAT", {"default": 0.0, "min": 0, "max": 1, "step": 0.05}), + }, + } + + RETURN_TYPES = ("IMAGE",) + FUNCTION = "prep_image" + + CATEGORY = "ipadapter" + + def prep_image(self, image, interpolation="LANCZOS", crop_position="center", sharpening=0.0): + size = (224, 224) + output = prepImage(image, interpolation, crop_position, size, sharpening, 0) + return (output, ) + +class IPAdapterEncoder: + @classmethod + def INPUT_TYPES(s): + return {"required": { + "clip_vision": ("CLIP_VISION",), + "image_1": ("IMAGE",), + "ipadapter_plus": ("BOOLEAN", { "default": False }), + "noise": ("FLOAT", { "default": 0.0, "min": 0.0, "max": 1.0, "step": 0.01 }), + "weight_1": ("FLOAT", { "default": 1.0, "min": 0, "max": 1.0, "step": 0.01 }), + }, + "optional": { + "image_2": ("IMAGE",), + "image_3": ("IMAGE",), + "image_4": ("IMAGE",), + "weight_2": ("FLOAT", { "default": 1.0, "min": 0, "max": 1.0, "step": 0.01 }), + "weight_3": ("FLOAT", { "default": 1.0, "min": 0, "max": 1.0, "step": 0.01 }), + "weight_4": ("FLOAT", { "default": 1.0, "min": 0, "max": 1.0, "step": 0.01 }), + } + } + + RETURN_TYPES = ("EMBEDS",) + FUNCTION = "preprocess" + CATEGORY = "ipadapter" + + def preprocess(self, clip_vision, image_1, ipadapter_plus, noise, weight_1, image_2=None, image_3=None, image_4=None, weight_2=1.0, weight_3=1.0, weight_4=1.0): + weight_1 *= (0.1 + (weight_1 - 0.1)) + weight_2 *= (0.1 + (weight_2 - 0.1)) + weight_3 *= (0.1 + (weight_3 - 0.1)) + weight_4 *= (0.1 + (weight_4 - 0.1)) + + image = image_1 + weight = [weight_1]*image_1.shape[0] + + if image_2 is not None: + if image_1.shape[1:] != image_2.shape[1:]: + image_2 = ldm_patched.modules.utils.common_upscale(image_2.movedim(-1,1), image.shape[2], image.shape[1], "bilinear", "center").movedim(1,-1) + image = torch.cat((image, image_2), dim=0) + weight += [weight_2]*image_2.shape[0] + if image_3 is not None: + if image.shape[1:] != image_3.shape[1:]: + image_3 = ldm_patched.modules.utils.common_upscale(image_3.movedim(-1,1), image.shape[2], image.shape[1], "bilinear", "center").movedim(1,-1) + image = torch.cat((image, image_3), dim=0) + weight += [weight_3]*image_3.shape[0] + if image_4 is not None: + if image.shape[1:] != image_4.shape[1:]: + image_4 = ldm_patched.modules.utils.common_upscale(image_4.movedim(-1,1), image.shape[2], image.shape[1], "bilinear", "center").movedim(1,-1) + image = torch.cat((image, image_4), dim=0) + weight += [weight_4]*image_4.shape[0] + + clip_embed = clip_vision.encode_image(image) + neg_image = image_add_noise(image, noise) if noise > 0 else None + + if ipadapter_plus: + clip_embed = clip_embed.penultimate_hidden_states + if noise > 0: + clip_embed_zeroed = clip_vision.encode_image(neg_image).penultimate_hidden_states + else: + clip_embed_zeroed = zeroed_hidden_states(clip_vision, image.shape[0]) + else: + clip_embed = clip_embed.image_embeds + if noise > 0: + clip_embed_zeroed = clip_vision.encode_image(neg_image).image_embeds + else: + clip_embed_zeroed = torch.zeros_like(clip_embed) + + if any(e != 1.0 for e in weight): + weight = torch.tensor(weight).unsqueeze(-1) if not ipadapter_plus else torch.tensor(weight).unsqueeze(-1).unsqueeze(-1) + clip_embed = clip_embed * weight + + output = torch.stack((clip_embed, clip_embed_zeroed)) + + return( output, ) + +class IPAdapterApplyEncoded(IPAdapterApply): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "ipadapter": ("IPADAPTER", ), + "embeds": ("EMBEDS",), + "model": ("MODEL", ), + "weight": ("FLOAT", { "default": 1.0, "min": -1, "max": 3, "step": 0.05 }), + "weight_type": (["original", "linear", "channel penalty"], ), + "start_at": ("FLOAT", { "default": 0.0, "min": 0.0, "max": 1.0, "step": 0.001 }), + "end_at": ("FLOAT", { "default": 1.0, "min": 0.0, "max": 1.0, "step": 0.001 }), + "unfold_batch": ("BOOLEAN", { "default": False }), + }, + "optional": { + "attn_mask": ("MASK",), + } + } + + +class IPAdapterBatchEmbeds: + @classmethod + def INPUT_TYPES(s): + return {"required": { + "embed1": ("EMBEDS",), + "embed2": ("EMBEDS",), + }} + + RETURN_TYPES = ("EMBEDS",) + FUNCTION = "batch" + CATEGORY = "ipadapter" + + def batch(self, embed1, embed2): + return (torch.cat((embed1, embed2), dim=1), ) + diff --git a/extensions-builtin/sd_forge_ipadapter/LICENSE b/extensions-builtin/sd_forge_ipadapter/LICENSE new file mode 100644 index 00000000..f288702d --- /dev/null +++ b/extensions-builtin/sd_forge_ipadapter/LICENSE @@ -0,0 +1,674 @@ + GNU GENERAL PUBLIC LICENSE + Version 3, 29 June 2007 + + Copyright (C) 2007 Free Software Foundation, Inc. + Everyone is permitted to copy and distribute verbatim copies + of this license document, but changing it is not allowed. + + Preamble + + The GNU General Public License is a free, copyleft license for +software and other kinds of works. + + The licenses for most software and other practical works are designed +to take away your freedom to share and change the works. By contrast, +the GNU General Public License is intended to guarantee your freedom to +share and change all versions of a program--to make sure it remains free +software for all its users. We, the Free Software Foundation, use the +GNU General Public License for most of our software; it applies also to +any other work released this way by its authors. You can apply it to +your programs, too. + + When we speak of free software, we are referring to freedom, not +price. Our General Public Licenses are designed to make sure that you +have the freedom to distribute copies of free software (and charge for +them if you wish), that you receive source code or can get it if you +want it, that you can change the software or use pieces of it in new +free programs, and that you know you can do these things. + + To protect your rights, we need to prevent others from denying you +these rights or asking you to surrender the rights. Therefore, you have +certain responsibilities if you distribute copies of the software, or if +you modify it: responsibilities to respect the freedom of others. + + For example, if you distribute copies of such a program, whether +gratis or for a fee, you must pass on to the recipients the same +freedoms that you received. You must make sure that they, too, receive +or can get the source code. And you must show them these terms so they +know their rights. + + Developers that use the GNU GPL protect your rights with two steps: +(1) assert copyright on the software, and (2) offer you this License +giving you legal permission to copy, distribute and/or modify it. + + For the developers' and authors' protection, the GPL clearly explains +that there is no warranty for this free software. For both users' and +authors' sake, the GPL requires that modified versions be marked as +changed, so that their problems will not be attributed erroneously to +authors of previous versions. + + Some devices are designed to deny users access to install or run +modified versions of the software inside them, although the manufacturer +can do so. This is fundamentally incompatible with the aim of +protecting users' freedom to change the software. The systematic +pattern of such abuse occurs in the area of products for individuals to +use, which is precisely where it is most unacceptable. Therefore, we +have designed this version of the GPL to prohibit the practice for those +products. If such problems arise substantially in other domains, we +stand ready to extend this provision to those domains in future versions +of the GPL, as needed to protect the freedom of users. + + Finally, every program is threatened constantly by software patents. +States should not allow patents to restrict development and use of +software on general-purpose computers, but in those that do, we wish to +avoid the special danger that patents applied to a free program could +make it effectively proprietary. To prevent this, the GPL assures that +patents cannot be used to render the program non-free. + + The precise terms and conditions for copying, distribution and +modification follow. + + TERMS AND CONDITIONS + + 0. Definitions. + + "This License" refers to version 3 of the GNU General Public License. + + "Copyright" also means copyright-like laws that apply to other kinds of +works, such as semiconductor masks. + + "The Program" refers to any copyrightable work licensed under this +License. Each licensee is addressed as "you". "Licensees" and +"recipients" may be individuals or organizations. + + To "modify" a work means to copy from or adapt all or part of the work +in a fashion requiring copyright permission, other than the making of an +exact copy. The resulting work is called a "modified version" of the +earlier work or a work "based on" the earlier work. + + A "covered work" means either the unmodified Program or a work based +on the Program. + + To "propagate" a work means to do anything with it that, without +permission, would make you directly or secondarily liable for +infringement under applicable copyright law, except executing it on a +computer or modifying a private copy. Propagation includes copying, +distribution (with or without modification), making available to the +public, and in some countries other activities as well. + + To "convey" a work means any kind of propagation that enables other +parties to make or receive copies. Mere interaction with a user through +a computer network, with no transfer of a copy, is not conveying. + + An interactive user interface displays "Appropriate Legal Notices" +to the extent that it includes a convenient and prominently visible +feature that (1) displays an appropriate copyright notice, and (2) +tells the user that there is no warranty for the work (except to the +extent that warranties are provided), that licensees may convey the +work under this License, and how to view a copy of this License. If +the interface presents a list of user commands or options, such as a +menu, a prominent item in the list meets this criterion. + + 1. Source Code. + + The "source code" for a work means the preferred form of the work +for making modifications to it. "Object code" means any non-source +form of a work. + + A "Standard Interface" means an interface that either is an official +standard defined by a recognized standards body, or, in the case of +interfaces specified for a particular programming language, one that +is widely used among developers working in that language. + + The "System Libraries" of an executable work include anything, other +than the work as a whole, that (a) is included in the normal form of +packaging a Major Component, but which is not part of that Major +Component, and (b) serves only to enable use of the work with that +Major Component, or to implement a Standard Interface for which an +implementation is available to the public in source code form. A +"Major Component", in this context, means a major essential component +(kernel, window system, and so on) of the specific operating system +(if any) on which the executable work runs, or a compiler used to +produce the work, or an object code interpreter used to run it. + + The "Corresponding Source" for a work in object code form means all +the source code needed to generate, install, and (for an executable +work) run the object code and to modify the work, including scripts to +control those activities. However, it does not include the work's +System Libraries, or general-purpose tools or generally available free +programs which are used unmodified in performing those activities but +which are not part of the work. For example, Corresponding Source +includes interface definition files associated with source files for +the work, and the source code for shared libraries and dynamically +linked subprograms that the work is specifically designed to require, +such as by intimate data communication or control flow between those +subprograms and other parts of the work. + + The Corresponding Source need not include anything that users +can regenerate automatically from other parts of the Corresponding +Source. + + The Corresponding Source for a work in source code form is that +same work. + + 2. Basic Permissions. + + All rights granted under this License are granted for the term of +copyright on the Program, and are irrevocable provided the stated +conditions are met. This License explicitly affirms your unlimited +permission to run the unmodified Program. The output from running a +covered work is covered by this License only if the output, given its +content, constitutes a covered work. This License acknowledges your +rights of fair use or other equivalent, as provided by copyright law. + + You may make, run and propagate covered works that you do not +convey, without conditions so long as your license otherwise remains +in force. You may convey covered works to others for the sole purpose +of having them make modifications exclusively for you, or provide you +with facilities for running those works, provided that you comply with +the terms of this License in conveying all material for which you do +not control copyright. Those thus making or running the covered works +for you must do so exclusively on your behalf, under your direction +and control, on terms that prohibit them from making any copies of +your copyrighted material outside their relationship with you. + + Conveying under any other circumstances is permitted solely under +the conditions stated below. Sublicensing is not allowed; section 10 +makes it unnecessary. + + 3. Protecting Users' Legal Rights From Anti-Circumvention Law. + + No covered work shall be deemed part of an effective technological +measure under any applicable law fulfilling obligations under article +11 of the WIPO copyright treaty adopted on 20 December 1996, or +similar laws prohibiting or restricting circumvention of such +measures. + + When you convey a covered work, you waive any legal power to forbid +circumvention of technological measures to the extent such circumvention +is effected by exercising rights under this License with respect to +the covered work, and you disclaim any intention to limit operation or +modification of the work as a means of enforcing, against the work's +users, your or third parties' legal rights to forbid circumvention of +technological measures. + + 4. Conveying Verbatim Copies. + + You may convey verbatim copies of the Program's source code as you +receive it, in any medium, provided that you conspicuously and +appropriately publish on each copy an appropriate copyright notice; +keep intact all notices stating that this License and any +non-permissive terms added in accord with section 7 apply to the code; +keep intact all notices of the absence of any warranty; and give all +recipients a copy of this License along with the Program. + + You may charge any price or no price for each copy that you convey, +and you may offer support or warranty protection for a fee. + + 5. Conveying Modified Source Versions. + + You may convey a work based on the Program, or the modifications to +produce it from the Program, in the form of source code under the +terms of section 4, provided that you also meet all of these conditions: + + a) The work must carry prominent notices stating that you modified + it, and giving a relevant date. + + b) The work must carry prominent notices stating that it is + released under this License and any conditions added under section + 7. This requirement modifies the requirement in section 4 to + "keep intact all notices". + + c) You must license the entire work, as a whole, under this + License to anyone who comes into possession of a copy. This + License will therefore apply, along with any applicable section 7 + additional terms, to the whole of the work, and all its parts, + regardless of how they are packaged. This License gives no + permission to license the work in any other way, but it does not + invalidate such permission if you have separately received it. + + d) If the work has interactive user interfaces, each must display + Appropriate Legal Notices; however, if the Program has interactive + interfaces that do not display Appropriate Legal Notices, your + work need not make them do so. + + A compilation of a covered work with other separate and independent +works, which are not by their nature extensions of the covered work, +and which are not combined with it such as to form a larger program, +in or on a volume of a storage or distribution medium, is called an +"aggregate" if the compilation and its resulting copyright are not +used to limit the access or legal rights of the compilation's users +beyond what the individual works permit. Inclusion of a covered work +in an aggregate does not cause this License to apply to the other +parts of the aggregate. + + 6. Conveying Non-Source Forms. + + You may convey a covered work in object code form under the terms +of sections 4 and 5, provided that you also convey the +machine-readable Corresponding Source under the terms of this License, +in one of these ways: + + a) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by the + Corresponding Source fixed on a durable physical medium + customarily used for software interchange. + + b) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by a + written offer, valid for at least three years and valid for as + long as you offer spare parts or customer support for that product + model, to give anyone who possesses the object code either (1) a + copy of the Corresponding Source for all the software in the + product that is covered by this License, on a durable physical + medium customarily used for software interchange, for a price no + more than your reasonable cost of physically performing this + conveying of source, or (2) access to copy the + Corresponding Source from a network server at no charge. + + c) Convey individual copies of the object code with a copy of the + written offer to provide the Corresponding Source. This + alternative is allowed only occasionally and noncommercially, and + only if you received the object code with such an offer, in accord + with subsection 6b. + + d) Convey the object code by offering access from a designated + place (gratis or for a charge), and offer equivalent access to the + Corresponding Source in the same way through the same place at no + further charge. You need not require recipients to copy the + Corresponding Source along with the object code. If the place to + copy the object code is a network server, the Corresponding Source + may be on a different server (operated by you or a third party) + that supports equivalent copying facilities, provided you maintain + clear directions next to the object code saying where to find the + Corresponding Source. Regardless of what server hosts the + Corresponding Source, you remain obligated to ensure that it is + available for as long as needed to satisfy these requirements. + + e) Convey the object code using peer-to-peer transmission, provided + you inform other peers where the object code and Corresponding + Source of the work are being offered to the general public at no + charge under subsection 6d. + + A separable portion of the object code, whose source code is excluded +from the Corresponding Source as a System Library, need not be +included in conveying the object code work. + + A "User Product" is either (1) a "consumer product", which means any +tangible personal property which is normally used for personal, family, +or household purposes, or (2) anything designed or sold for incorporation +into a dwelling. In determining whether a product is a consumer product, +doubtful cases shall be resolved in favor of coverage. For a particular +product received by a particular user, "normally used" refers to a +typical or common use of that class of product, regardless of the status +of the particular user or of the way in which the particular user +actually uses, or expects or is expected to use, the product. A product +is a consumer product regardless of whether the product has substantial +commercial, industrial or non-consumer uses, unless such uses represent +the only significant mode of use of the product. + + "Installation Information" for a User Product means any methods, +procedures, authorization keys, or other information required to install +and execute modified versions of a covered work in that User Product from +a modified version of its Corresponding Source. The information must +suffice to ensure that the continued functioning of the modified object +code is in no case prevented or interfered with solely because +modification has been made. + + If you convey an object code work under this section in, or with, or +specifically for use in, a User Product, and the conveying occurs as +part of a transaction in which the right of possession and use of the +User Product is transferred to the recipient in perpetuity or for a +fixed term (regardless of how the transaction is characterized), the +Corresponding Source conveyed under this section must be accompanied +by the Installation Information. But this requirement does not apply +if neither you nor any third party retains the ability to install +modified object code on the User Product (for example, the work has +been installed in ROM). + + The requirement to provide Installation Information does not include a +requirement to continue to provide support service, warranty, or updates +for a work that has been modified or installed by the recipient, or for +the User Product in which it has been modified or installed. Access to a +network may be denied when the modification itself materially and +adversely affects the operation of the network or violates the rules and +protocols for communication across the network. + + Corresponding Source conveyed, and Installation Information provided, +in accord with this section must be in a format that is publicly +documented (and with an implementation available to the public in +source code form), and must require no special password or key for +unpacking, reading or copying. + + 7. Additional Terms. + + "Additional permissions" are terms that supplement the terms of this +License by making exceptions from one or more of its conditions. +Additional permissions that are applicable to the entire Program shall +be treated as though they were included in this License, to the extent +that they are valid under applicable law. If additional permissions +apply only to part of the Program, that part may be used separately +under those permissions, but the entire Program remains governed by +this License without regard to the additional permissions. + + When you convey a copy of a covered work, you may at your option +remove any additional permissions from that copy, or from any part of +it. (Additional permissions may be written to require their own +removal in certain cases when you modify the work.) You may place +additional permissions on material, added by you to a covered work, +for which you have or can give appropriate copyright permission. + + Notwithstanding any other provision of this License, for material you +add to a covered work, you may (if authorized by the copyright holders of +that material) supplement the terms of this License with terms: + + a) Disclaiming warranty or limiting liability differently from the + terms of sections 15 and 16 of this License; or + + b) Requiring preservation of specified reasonable legal notices or + author attributions in that material or in the Appropriate Legal + Notices displayed by works containing it; or + + c) Prohibiting misrepresentation of the origin of that material, or + requiring that modified versions of such material be marked in + reasonable ways as different from the original version; or + + d) Limiting the use for publicity purposes of names of licensors or + authors of the material; or + + e) Declining to grant rights under trademark law for use of some + trade names, trademarks, or service marks; or + + f) Requiring indemnification of licensors and authors of that + material by anyone who conveys the material (or modified versions of + it) with contractual assumptions of liability to the recipient, for + any liability that these contractual assumptions directly impose on + those licensors and authors. + + All other non-permissive additional terms are considered "further +restrictions" within the meaning of section 10. If the Program as you +received it, or any part of it, contains a notice stating that it is +governed by this License along with a term that is a further +restriction, you may remove that term. If a license document contains +a further restriction but permits relicensing or conveying under this +License, you may add to a covered work material governed by the terms +of that license document, provided that the further restriction does +not survive such relicensing or conveying. + + If you add terms to a covered work in accord with this section, you +must place, in the relevant source files, a statement of the +additional terms that apply to those files, or a notice indicating +where to find the applicable terms. + + Additional terms, permissive or non-permissive, may be stated in the +form of a separately written license, or stated as exceptions; +the above requirements apply either way. + + 8. Termination. + + You may not propagate or modify a covered work except as expressly +provided under this License. Any attempt otherwise to propagate or +modify it is void, and will automatically terminate your rights under +this License (including any patent licenses granted under the third +paragraph of section 11). + + However, if you cease all violation of this License, then your +license from a particular copyright holder is reinstated (a) +provisionally, unless and until the copyright holder explicitly and +finally terminates your license, and (b) permanently, if the copyright +holder fails to notify you of the violation by some reasonable means +prior to 60 days after the cessation. + + Moreover, your license from a particular copyright holder is +reinstated permanently if the copyright holder notifies you of the +violation by some reasonable means, this is the first time you have +received notice of violation of this License (for any work) from that +copyright holder, and you cure the violation prior to 30 days after +your receipt of the notice. + + Termination of your rights under this section does not terminate the +licenses of parties who have received copies or rights from you under +this License. If your rights have been terminated and not permanently +reinstated, you do not qualify to receive new licenses for the same +material under section 10. + + 9. Acceptance Not Required for Having Copies. + + You are not required to accept this License in order to receive or +run a copy of the Program. Ancillary propagation of a covered work +occurring solely as a consequence of using peer-to-peer transmission +to receive a copy likewise does not require acceptance. However, +nothing other than this License grants you permission to propagate or +modify any covered work. These actions infringe copyright if you do +not accept this License. Therefore, by modifying or propagating a +covered work, you indicate your acceptance of this License to do so. + + 10. Automatic Licensing of Downstream Recipients. + + Each time you convey a covered work, the recipient automatically +receives a license from the original licensors, to run, modify and +propagate that work, subject to this License. You are not responsible +for enforcing compliance by third parties with this License. + + An "entity transaction" is a transaction transferring control of an +organization, or substantially all assets of one, or subdividing an +organization, or merging organizations. If propagation of a covered +work results from an entity transaction, each party to that +transaction who receives a copy of the work also receives whatever +licenses to the work the party's predecessor in interest had or could +give under the previous paragraph, plus a right to possession of the +Corresponding Source of the work from the predecessor in interest, if +the predecessor has it or can get it with reasonable efforts. + + You may not impose any further restrictions on the exercise of the +rights granted or affirmed under this License. For example, you may +not impose a license fee, royalty, or other charge for exercise of +rights granted under this License, and you may not initiate litigation +(including a cross-claim or counterclaim in a lawsuit) alleging that +any patent claim is infringed by making, using, selling, offering for +sale, or importing the Program or any portion of it. + + 11. Patents. + + A "contributor" is a copyright holder who authorizes use under this +License of the Program or a work on which the Program is based. The +work thus licensed is called the contributor's "contributor version". + + A contributor's "essential patent claims" are all patent claims +owned or controlled by the contributor, whether already acquired or +hereafter acquired, that would be infringed by some manner, permitted +by this License, of making, using, or selling its contributor version, +but do not include claims that would be infringed only as a +consequence of further modification of the contributor version. For +purposes of this definition, "control" includes the right to grant +patent sublicenses in a manner consistent with the requirements of +this License. + + Each contributor grants you a non-exclusive, worldwide, royalty-free +patent license under the contributor's essential patent claims, to +make, use, sell, offer for sale, import and otherwise run, modify and +propagate the contents of its contributor version. + + In the following three paragraphs, a "patent license" is any express +agreement or commitment, however denominated, not to enforce a patent +(such as an express permission to practice a patent or covenant not to +sue for patent infringement). To "grant" such a patent license to a +party means to make such an agreement or commitment not to enforce a +patent against the party. + + If you convey a covered work, knowingly relying on a patent license, +and the Corresponding Source of the work is not available for anyone +to copy, free of charge and under the terms of this License, through a +publicly available network server or other readily accessible means, +then you must either (1) cause the Corresponding Source to be so +available, or (2) arrange to deprive yourself of the benefit of the +patent license for this particular work, or (3) arrange, in a manner +consistent with the requirements of this License, to extend the patent +license to downstream recipients. "Knowingly relying" means you have +actual knowledge that, but for the patent license, your conveying the +covered work in a country, or your recipient's use of the covered work +in a country, would infringe one or more identifiable patents in that +country that you have reason to believe are valid. + + If, pursuant to or in connection with a single transaction or +arrangement, you convey, or propagate by procuring conveyance of, a +covered work, and grant a patent license to some of the parties +receiving the covered work authorizing them to use, propagate, modify +or convey a specific copy of the covered work, then the patent license +you grant is automatically extended to all recipients of the covered +work and works based on it. + + A patent license is "discriminatory" if it does not include within +the scope of its coverage, prohibits the exercise of, or is +conditioned on the non-exercise of one or more of the rights that are +specifically granted under this License. You may not convey a covered +work if you are a party to an arrangement with a third party that is +in the business of distributing software, under which you make payment +to the third party based on the extent of your activity of conveying +the work, and under which the third party grants, to any of the +parties who would receive the covered work from you, a discriminatory +patent license (a) in connection with copies of the covered work +conveyed by you (or copies made from those copies), or (b) primarily +for and in connection with specific products or compilations that +contain the covered work, unless you entered into that arrangement, +or that patent license was granted, prior to 28 March 2007. + + Nothing in this License shall be construed as excluding or limiting +any implied license or other defenses to infringement that may +otherwise be available to you under applicable patent law. + + 12. No Surrender of Others' Freedom. + + If conditions are imposed on you (whether by court order, agreement or +otherwise) that contradict the conditions of this License, they do not +excuse you from the conditions of this License. If you cannot convey a +covered work so as to satisfy simultaneously your obligations under this +License and any other pertinent obligations, then as a consequence you may +not convey it at all. For example, if you agree to terms that obligate you +to collect a royalty for further conveying from those to whom you convey +the Program, the only way you could satisfy both those terms and this +License would be to refrain entirely from conveying the Program. + + 13. Use with the GNU Affero General Public License. + + Notwithstanding any other provision of this License, you have +permission to link or combine any covered work with a work licensed +under version 3 of the GNU Affero General Public License into a single +combined work, and to convey the resulting work. The terms of this +License will continue to apply to the part which is the covered work, +but the special requirements of the GNU Affero General Public License, +section 13, concerning interaction through a network will apply to the +combination as such. + + 14. Revised Versions of this License. + + The Free Software Foundation may publish revised and/or new versions of +the GNU General Public License from time to time. Such new versions will +be similar in spirit to the present version, but may differ in detail to +address new problems or concerns. + + Each version is given a distinguishing version number. If the +Program specifies that a certain numbered version of the GNU General +Public License "or any later version" applies to it, you have the +option of following the terms and conditions either of that numbered +version or of any later version published by the Free Software +Foundation. If the Program does not specify a version number of the +GNU General Public License, you may choose any version ever published +by the Free Software Foundation. + + If the Program specifies that a proxy can decide which future +versions of the GNU General Public License can be used, that proxy's +public statement of acceptance of a version permanently authorizes you +to choose that version for the Program. + + Later license versions may give you additional or different +permissions. However, no additional obligations are imposed on any +author or copyright holder as a result of your choosing to follow a +later version. + + 15. Disclaimer of Warranty. + + THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY +APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT +HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY +OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, +THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR +PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM +IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF +ALL NECESSARY SERVICING, REPAIR OR CORRECTION. + + 16. Limitation of Liability. + + IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING +WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS +THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY +GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE +USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF +DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD +PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), +EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF +SUCH DAMAGES. + + 17. Interpretation of Sections 15 and 16. + + If the disclaimer of warranty and limitation of liability provided +above cannot be given local legal effect according to their terms, +reviewing courts shall apply local law that most closely approximates +an absolute waiver of all civil liability in connection with the +Program, unless a warranty or assumption of liability accompanies a +copy of the Program in return for a fee. + + END OF TERMS AND CONDITIONS + + How to Apply These Terms to Your New Programs + + If you develop a new program, and you want it to be of the greatest +possible use to the public, the best way to achieve this is to make it +free software which everyone can redistribute and change under these terms. + + To do so, attach the following notices to the program. It is safest +to attach them to the start of each source file to most effectively +state the exclusion of warranty; and each file should have at least +the "copyright" line and a pointer to where the full notice is found. + + + Copyright (C) + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . + +Also add information on how to contact you by electronic and paper mail. + + If the program does terminal interaction, make it output a short +notice like this when it starts in an interactive mode: + + Copyright (C) + This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. + This is free software, and you are welcome to redistribute it + under certain conditions; type `show c' for details. + +The hypothetical commands `show w' and `show c' should show the appropriate +parts of the General Public License. Of course, your program's commands +might be different; for a GUI interface, you would use an "about box". + + You should also get your employer (if you work as a programmer) or school, +if any, to sign a "copyright disclaimer" for the program, if necessary. +For more information on this, and how to apply and follow the GNU GPL, see +. + + The GNU General Public License does not permit incorporating your program +into proprietary programs. If your program is a subroutine library, you +may consider it more useful to permit linking proprietary applications with +the library. If this is what you want to do, use the GNU Lesser General +Public License instead of this License. But first, please read +. diff --git a/extensions-builtin/sd_forge_ipadapter/resampler.py b/extensions-builtin/sd_forge_ipadapter/resampler.py new file mode 100644 index 00000000..24266671 --- /dev/null +++ b/extensions-builtin/sd_forge_ipadapter/resampler.py @@ -0,0 +1,158 @@ +# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py +# and https://github.com/lucidrains/imagen-pytorch/blob/main/imagen_pytorch/imagen_pytorch.py + +import math + +import torch +import torch.nn as nn +from einops import rearrange +from einops.layers.torch import Rearrange + + +# FFN +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.pos_emb = nn.Embedding(max_seq_len, embedding_dim) if apply_pos_emb else None + + 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.to_latents_from_mean_pooled_seq = ( + nn.Sequential( + nn.LayerNorm(dim), + nn.Linear(dim, dim * num_latents_mean_pooled), + Rearrange("b (n d) -> b n d", n=num_latents_mean_pooled), + ) + if num_latents_mean_pooled > 0 + else None + ) + + 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): + if self.pos_emb is not None: + n, device = x.shape[1], x.device + pos_emb = self.pos_emb(torch.arange(n, device=device)) + x = x + pos_emb + + latents = self.latents.repeat(x.size(0), 1, 1) + + x = self.proj_in(x) + + if self.to_latents_from_mean_pooled_seq: + meanpooled_seq = masked_mean(x, dim=1, mask=torch.ones(x.shape[:2], device=x.device, dtype=torch.bool)) + meanpooled_latents = self.to_latents_from_mean_pooled_seq(meanpooled_seq) + latents = torch.cat((meanpooled_latents, latents), dim=-2) + + 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 masked_mean(t, *, dim, mask=None): + if mask is None: + return t.mean(dim=dim) + + denom = mask.sum(dim=dim, keepdim=True) + mask = rearrange(mask, "b n -> b n 1") + masked_t = t.masked_fill(~mask, 0.0) + + return masked_t.sum(dim=dim) / denom.clamp(min=1e-5) diff --git a/extensions-builtin/sd_forge_ipadapter/thanks b/extensions-builtin/sd_forge_ipadapter/thanks new file mode 100644 index 00000000..e1c7bcb4 --- /dev/null +++ b/extensions-builtin/sd_forge_ipadapter/thanks @@ -0,0 +1 @@ +This repo is modified from https://github.com/cubiq/ComfyUI_IPAdapter_plus