Merge branch 'master' into worksplit-multigpu

comfy/latent_formats.py

@@ -456,3 +456,13 @@ class Wan21(LatentFormat):
         latents_mean = self.latents_mean.to(latent.device, latent.dtype)
         latents_std = self.latents_std.to(latent.device, latent.dtype)
         return latent * latents_std / self.scale_factor + latents_mean
+
+class Hunyuan3Dv2(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 1
+    scale_factor = 0.9990943042622529
+
+class Hunyuan3Dv2mini(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 1
+    scale_factor = 1.0188137142395404

comfy/ldm/hunyuan3d/model.py

@@ -0,0 +1,135 @@
+import torch
+from torch import nn
+from comfy.ldm.flux.layers import (
+    DoubleStreamBlock,
+    LastLayer,
+    MLPEmbedder,
+    SingleStreamBlock,
+    timestep_embedding,
+)
+
+
+class Hunyuan3Dv2(nn.Module):
+    def __init__(
+        self,
+        in_channels=64,
+        context_in_dim=1536,
+        hidden_size=1024,
+        mlp_ratio=4.0,
+        num_heads=16,
+        depth=16,
+        depth_single_blocks=32,
+        qkv_bias=True,
+        guidance_embed=False,
+        image_model=None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.dtype = dtype
+
+        if hidden_size % num_heads != 0:
+            raise ValueError(
+                f"Hidden size {hidden_size} must be divisible by num_heads {num_heads}"
+            )
+
+        self.max_period = 1000  # While reimplementing the model I noticed that they messed up. This 1000 value was meant to be the time_factor but they set the max_period instead
+        self.latent_in = operations.Linear(in_channels, hidden_size, bias=True, dtype=dtype, device=device)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=hidden_size, dtype=dtype, device=device, operations=operations)
+        self.guidance_in = (
+            MLPEmbedder(in_dim=256, hidden_dim=hidden_size, dtype=dtype, device=device, operations=operations) if guidance_embed else None
+        )
+        self.cond_in = operations.Linear(context_in_dim, hidden_size, dtype=dtype, device=device)
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(depth)
+            ]
+        )
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(depth_single_blocks)
+            ]
+        )
+        self.final_layer = LastLayer(hidden_size, 1, in_channels, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, x, timestep, context, guidance=None, transformer_options={}, **kwargs):
+        x = x.movedim(-1, -2)
+        timestep = 1.0 - timestep
+        txt = context
+        img = self.latent_in(x)
+
+        vec = self.time_in(timestep_embedding(timestep, 256, self.max_period).to(dtype=img.dtype))
+        if self.guidance_in is not None:
+            if guidance is not None:
+                vec = vec + self.guidance_in(timestep_embedding(guidance, 256, self.max_period).to(img.dtype))
+
+        txt = self.cond_in(txt)
+        pe = None
+        attn_mask = None
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.double_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"], out["txt"] = block(img=args["img"],
+                                                   txt=args["txt"],
+                                                   vec=args["vec"],
+                                                   pe=args["pe"],
+                                                   attn_mask=args.get("attn_mask"))
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": img,
+                                                           "txt": txt,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask},
+                                                          {"original_block": block_wrap})
+                txt = out["txt"]
+                img = out["img"]
+            else:
+                img, txt = block(img=img,
+                                 txt=txt,
+                                 vec=vec,
+                                 pe=pe,
+                                 attn_mask=attn_mask)
+
+        img = torch.cat((txt, img), 1)
+
+        for i, block in enumerate(self.single_blocks):
+            if ("single_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"],
+                                       vec=args["vec"],
+                                       pe=args["pe"],
+                                       attn_mask=args.get("attn_mask"))
+                    return out
+
+                out = blocks_replace[("single_block", i)]({"img": img,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask},
+                                                          {"original_block": block_wrap})
+                img = out["img"]
+            else:
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
+
+        img = img[:, txt.shape[1]:, ...]
+        img = self.final_layer(img, vec)
+        return img.movedim(-2, -1) * (-1.0)

comfy/ldm/hunyuan3d/vae.py

@@ -0,0 +1,587 @@
+# Original: https://github.com/Tencent/Hunyuan3D-2/blob/main/hy3dgen/shapegen/models/autoencoders/model.py
+# Since the header on their VAE source file was a bit confusing we asked for permission to use this code from tencent under the GPL license used in ComfyUI.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+from typing import Union, Tuple, List, Callable, Optional
+
+import numpy as np
+from einops import repeat, rearrange
+from tqdm import tqdm
+import logging
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+def generate_dense_grid_points(
+    bbox_min: np.ndarray,
+    bbox_max: np.ndarray,
+    octree_resolution: int,
+    indexing: str = "ij",
+):
+    length = bbox_max - bbox_min
+    num_cells = octree_resolution
+
+    x = np.linspace(bbox_min[0], bbox_max[0], int(num_cells) + 1, dtype=np.float32)
+    y = np.linspace(bbox_min[1], bbox_max[1], int(num_cells) + 1, dtype=np.float32)
+    z = np.linspace(bbox_min[2], bbox_max[2], int(num_cells) + 1, dtype=np.float32)
+    [xs, ys, zs] = np.meshgrid(x, y, z, indexing=indexing)
+    xyz = np.stack((xs, ys, zs), axis=-1)
+    grid_size = [int(num_cells) + 1, int(num_cells) + 1, int(num_cells) + 1]
+
+    return xyz, grid_size, length
+
+
+class VanillaVolumeDecoder:
+    @torch.no_grad()
+    def __call__(
+        self,
+        latents: torch.FloatTensor,
+        geo_decoder: Callable,
+        bounds: Union[Tuple[float], List[float], float] = 1.01,
+        num_chunks: int = 10000,
+        octree_resolution: int = None,
+        enable_pbar: bool = True,
+        **kwargs,
+    ):
+        device = latents.device
+        dtype = latents.dtype
+        batch_size = latents.shape[0]
+
+        # 1. generate query points
+        if isinstance(bounds, float):
+            bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds]
+
+        bbox_min, bbox_max = np.array(bounds[0:3]), np.array(bounds[3:6])
+        xyz_samples, grid_size, length = generate_dense_grid_points(
+            bbox_min=bbox_min,
+            bbox_max=bbox_max,
+            octree_resolution=octree_resolution,
+            indexing="ij"
+        )
+        xyz_samples = torch.from_numpy(xyz_samples).to(device, dtype=dtype).contiguous().reshape(-1, 3)
+
+        # 2. latents to 3d volume
+        batch_logits = []
+        for start in tqdm(range(0, xyz_samples.shape[0], num_chunks), desc="Volume Decoding",
+                          disable=not enable_pbar):
+            chunk_queries = xyz_samples[start: start + num_chunks, :]
+            chunk_queries = repeat(chunk_queries, "p c -> b p c", b=batch_size)
+            logits = geo_decoder(queries=chunk_queries, latents=latents)
+            batch_logits.append(logits)
+
+        grid_logits = torch.cat(batch_logits, dim=1)
+        grid_logits = grid_logits.view((batch_size, *grid_size)).float()
+
+        return grid_logits
+
+
+class FourierEmbedder(nn.Module):
+    """The sin/cosine positional embedding. Given an input tensor `x` of shape [n_batch, ..., c_dim], it converts
+    each feature dimension of `x[..., i]` into:
+        [
+            sin(x[..., i]),
+            sin(f_1*x[..., i]),
+            sin(f_2*x[..., i]),
+            ...
+            sin(f_N * x[..., i]),
+            cos(x[..., i]),
+            cos(f_1*x[..., i]),
+            cos(f_2*x[..., i]),
+            ...
+            cos(f_N * x[..., i]),
+            x[..., i]     # only present if include_input is True.
+        ], here f_i is the frequency.
+
+    Denote the space is [0 / num_freqs, 1 / num_freqs, 2 / num_freqs, 3 / num_freqs, ..., (num_freqs - 1) / num_freqs].
+    If logspace is True, then the frequency f_i is [2^(0 / num_freqs), ..., 2^(i / num_freqs), ...];
+    Otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)].
+
+    Args:
+        num_freqs (int): the number of frequencies, default is 6;
+        logspace (bool): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+            otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)];
+        input_dim (int): the input dimension, default is 3;
+        include_input (bool): include the input tensor or not, default is True.
+
+    Attributes:
+        frequencies (torch.Tensor): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+                otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1);
+
+        out_dim (int): the embedding size, if include_input is True, it is input_dim * (num_freqs * 2 + 1),
+            otherwise, it is input_dim * num_freqs * 2.
+
+    """
+
+    def __init__(self,
+                 num_freqs: int = 6,
+                 logspace: bool = True,
+                 input_dim: int = 3,
+                 include_input: bool = True,
+                 include_pi: bool = True) -> None:
+
+        """The initialization"""
+
+        super().__init__()
+
+        if logspace:
+            frequencies = 2.0 ** torch.arange(
+                num_freqs,
+                dtype=torch.float32
+            )
+        else:
+            frequencies = torch.linspace(
+                1.0,
+                2.0 ** (num_freqs - 1),
+                num_freqs,
+                dtype=torch.float32
+            )
+
+        if include_pi:
+            frequencies *= torch.pi
+
+        self.register_buffer("frequencies", frequencies, persistent=False)
+        self.include_input = include_input
+        self.num_freqs = num_freqs
+
+        self.out_dim = self.get_dims(input_dim)
+
+    def get_dims(self, input_dim):
+        temp = 1 if self.include_input or self.num_freqs == 0 else 0
+        out_dim = input_dim * (self.num_freqs * 2 + temp)
+
+        return out_dim
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """ Forward process.
+
+        Args:
+            x: tensor of shape [..., dim]
+
+        Returns:
+            embedding: an embedding of `x` of shape [..., dim * (num_freqs * 2 + temp)]
+                where temp is 1 if include_input is True and 0 otherwise.
+        """
+
+        if self.num_freqs > 0:
+            embed = (x[..., None].contiguous() * self.frequencies.to(device=x.device, dtype=x.dtype)).view(*x.shape[:-1], -1)
+            if self.include_input:
+                return torch.cat((x, embed.sin(), embed.cos()), dim=-1)
+            else:
+                return torch.cat((embed.sin(), embed.cos()), dim=-1)
+        else:
+            return x
+
+
+class CrossAttentionProcessor:
+    def __call__(self, attn, q, k, v):
+        out = F.scaled_dot_product_attention(q, k, v)
+        return out
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+
+    def __init__(self, drop_prob: float = 0., scale_by_keep: bool = True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+        This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+        the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+        See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+        changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+        'survival rate' as the argument.
+
+        """
+        if self.drop_prob == 0. or not self.training:
+            return x
+        keep_prob = 1 - self.drop_prob
+        shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+        random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+        if keep_prob > 0.0 and self.scale_by_keep:
+            random_tensor.div_(keep_prob)
+        return x * random_tensor
+
+    def extra_repr(self):
+        return f'drop_prob={round(self.drop_prob, 3):0.3f}'
+
+
+class MLP(nn.Module):
+    def __init__(
+        self, *,
+        width: int,
+        expand_ratio: int = 4,
+        output_width: int = None,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.c_fc = ops.Linear(width, width * expand_ratio)
+        self.c_proj = ops.Linear(width * expand_ratio, output_width if output_width is not None else width)
+        self.gelu = nn.GELU()
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+
+    def forward(self, x):
+        return self.drop_path(self.c_proj(self.gelu(self.c_fc(x))))
+
+
+class QKVMultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        heads: int,
+        width=None,
+        qk_norm=False,
+        norm_layer=ops.LayerNorm
+    ):
+        super().__init__()
+        self.heads = heads
+        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+
+        self.attn_processor = CrossAttentionProcessor()
+
+    def forward(self, q, kv):
+        _, n_ctx, _ = q.shape
+        bs, n_data, width = kv.shape
+        attn_ch = width // self.heads // 2
+        q = q.view(bs, n_ctx, self.heads, -1)
+        kv = kv.view(bs, n_data, self.heads, -1)
+        k, v = torch.split(kv, attn_ch, dim=-1)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
+        out = self.attn_processor(self, q, k, v)
+        out = out.transpose(1, 2).reshape(bs, n_ctx, -1)
+        return out
+
+
+class MultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        data_width: Optional[int] = None,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        kv_cache: bool = False,
+    ):
+        super().__init__()
+        self.width = width
+        self.heads = heads
+        self.data_width = width if data_width is None else data_width
+        self.c_q = ops.Linear(width, width, bias=qkv_bias)
+        self.c_kv = ops.Linear(self.data_width, width * 2, bias=qkv_bias)
+        self.c_proj = ops.Linear(width, width)
+        self.attention = QKVMultiheadCrossAttention(
+            heads=heads,
+            width=width,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.kv_cache = kv_cache
+        self.data = None
+
+    def forward(self, x, data):
+        x = self.c_q(x)
+        if self.kv_cache:
+            if self.data is None:
+                self.data = self.c_kv(data)
+                logging.info('Save kv cache,this should be called only once for one mesh')
+            data = self.data
+        else:
+            data = self.c_kv(data)
+        x = self.attention(x, data)
+        x = self.c_proj(x)
+        return x
+
+
+class ResidualCrossAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        mlp_expand_ratio: int = 4,
+        data_width: Optional[int] = None,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False
+    ):
+        super().__init__()
+
+        if data_width is None:
+            data_width = width
+
+        self.attn = MultiheadCrossAttention(
+            width=width,
+            heads=heads,
+            data_width=data_width,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.ln_2 = norm_layer(data_width, elementwise_affine=True, eps=1e-6)
+        self.ln_3 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.mlp = MLP(width=width, expand_ratio=mlp_expand_ratio)
+
+    def forward(self, x: torch.Tensor, data: torch.Tensor):
+        x = x + self.attn(self.ln_1(x), self.ln_2(data))
+        x = x + self.mlp(self.ln_3(x))
+        return x
+
+
+class QKVMultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        heads: int,
+        width=None,
+        qk_norm=False,
+        norm_layer=ops.LayerNorm
+    ):
+        super().__init__()
+        self.heads = heads
+        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+
+    def forward(self, qkv):
+        bs, n_ctx, width = qkv.shape
+        attn_ch = width // self.heads // 3
+        qkv = qkv.view(bs, n_ctx, self.heads, -1)
+        q, k, v = torch.split(qkv, attn_ch, dim=-1)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
+        out = F.scaled_dot_product_attention(q, k, v).transpose(1, 2).reshape(bs, n_ctx, -1)
+        return out
+
+
+class MultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.heads = heads
+        self.c_qkv = ops.Linear(width, width * 3, bias=qkv_bias)
+        self.c_proj = ops.Linear(width, width)
+        self.attention = QKVMultiheadAttention(
+            heads=heads,
+            width=width,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+
+    def forward(self, x):
+        x = self.c_qkv(x)
+        x = self.attention(x)
+        x = self.drop_path(self.c_proj(x))
+        return x
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0,
+    ):
+        super().__init__()
+        self.attn = MultiheadAttention(
+            width=width,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm,
+            drop_path_rate=drop_path_rate
+        )
+        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.mlp = MLP(width=width, drop_path_rate=drop_path_rate)
+        self.ln_2 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        layers: int,
+        heads: int,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    width=width,
+                    heads=heads,
+                    qkv_bias=qkv_bias,
+                    norm_layer=norm_layer,
+                    qk_norm=qk_norm,
+                    drop_path_rate=drop_path_rate
+                )
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor):
+        for block in self.resblocks:
+            x = block(x)
+        return x
+
+
+class CrossAttentionDecoder(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        out_channels: int,
+        fourier_embedder: FourierEmbedder,
+        width: int,
+        heads: int,
+        mlp_expand_ratio: int = 4,
+        downsample_ratio: int = 1,
+        enable_ln_post: bool = True,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        label_type: str = "binary"
+    ):
+        super().__init__()
+
+        self.enable_ln_post = enable_ln_post
+        self.fourier_embedder = fourier_embedder
+        self.downsample_ratio = downsample_ratio
+        self.query_proj = ops.Linear(self.fourier_embedder.out_dim, width)
+        if self.downsample_ratio != 1:
+            self.latents_proj = ops.Linear(width * downsample_ratio, width)
+        if self.enable_ln_post == False:
+            qk_norm = False
+        self.cross_attn_decoder = ResidualCrossAttentionBlock(
+            width=width,
+            mlp_expand_ratio=mlp_expand_ratio,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm
+        )
+
+        if self.enable_ln_post:
+            self.ln_post = ops.LayerNorm(width)
+        self.output_proj = ops.Linear(width, out_channels)
+        self.label_type = label_type
+        self.count = 0
+
+    def forward(self, queries=None, query_embeddings=None, latents=None):
+        if query_embeddings is None:
+            query_embeddings = self.query_proj(self.fourier_embedder(queries).to(latents.dtype))
+        self.count += query_embeddings.shape[1]
+        if self.downsample_ratio != 1:
+            latents = self.latents_proj(latents)
+        x = self.cross_attn_decoder(query_embeddings, latents)
+        if self.enable_ln_post:
+            x = self.ln_post(x)
+        occ = self.output_proj(x)
+        return occ
+
+
+class ShapeVAE(nn.Module):
+    def __init__(
+        self,
+        *,
+        embed_dim: int,
+        width: int,
+        heads: int,
+        num_decoder_layers: int,
+        geo_decoder_downsample_ratio: int = 1,
+        geo_decoder_mlp_expand_ratio: int = 4,
+        geo_decoder_ln_post: bool = True,
+        num_freqs: int = 8,
+        include_pi: bool = True,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        label_type: str = "binary",
+        drop_path_rate: float = 0.0,
+        scale_factor: float = 1.0,
+    ):
+        super().__init__()
+        self.geo_decoder_ln_post = geo_decoder_ln_post
+
+        self.fourier_embedder = FourierEmbedder(num_freqs=num_freqs, include_pi=include_pi)
+
+        self.post_kl = ops.Linear(embed_dim, width)
+
+        self.transformer = Transformer(
+            width=width,
+            layers=num_decoder_layers,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
+            drop_path_rate=drop_path_rate
+        )
+
+        self.geo_decoder = CrossAttentionDecoder(
+            fourier_embedder=self.fourier_embedder,
+            out_channels=1,
+            mlp_expand_ratio=geo_decoder_mlp_expand_ratio,
+            downsample_ratio=geo_decoder_downsample_ratio,
+            enable_ln_post=self.geo_decoder_ln_post,
+            width=width // geo_decoder_downsample_ratio,
+            heads=heads // geo_decoder_downsample_ratio,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
+            label_type=label_type,
+        )
+
+        self.volume_decoder = VanillaVolumeDecoder()
+        self.scale_factor = scale_factor
+
+    def decode(self, latents, **kwargs):
+        latents = self.post_kl(latents.movedim(-2, -1))
+        latents = self.transformer(latents)
+
+        bounds = kwargs.get("bounds", 1.01)
+        num_chunks = kwargs.get("num_chunks", 8000)
+        octree_resolution = kwargs.get("octree_resolution", 256)
+        enable_pbar = kwargs.get("enable_pbar", True)
+
+        grid_logits = self.volume_decoder(latents, self.geo_decoder, bounds=bounds, num_chunks=num_chunks, octree_resolution=octree_resolution, enable_pbar=enable_pbar)
+        return grid_logits.movedim(-2, -1)
+
+    def encode(self, x):
+        return None

comfy/ldm/modules/attention.py

@@ -471,7 +471,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
 def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
     if skip_reshape:
         b, _, _, dim_head = q.shape
-        tensor_layout="HND"
+        tensor_layout = "HND"
     else:
         b, _, dim_head = q.shape
         dim_head //= heads
@@ -479,7 +479,7 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
             lambda t: t.view(b, -1, heads, dim_head),
             (q, k, v),
         )
-        tensor_layout="NHD"
+        tensor_layout = "NHD"
 
     if mask is not None:
         # add a batch dimension if there isn't already one
@@ -489,7 +489,17 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
         if mask.ndim == 3:
             mask = mask.unsqueeze(1)
 
-    out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
+    try:
+        out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
+    except Exception as e:
+        logging.error("Error running sage attention: {}, using pytorch attention instead.".format(e))
+        if tensor_layout == "NHD":
+            q, k, v = map(
+                lambda t: t.transpose(1, 2),
+                (q, k, v),
+            )
+        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=True, skip_output_reshape=skip_output_reshape)
+
     if tensor_layout == "HND":
         if not skip_output_reshape:
             out = (

comfy/model_base.py

@@ -36,6 +36,7 @@ import comfy.ldm.hunyuan_video.model
 import comfy.ldm.cosmos.model
 import comfy.ldm.lumina.model
 import comfy.ldm.wan.model
+import comfy.ldm.hunyuan3d.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -58,6 +59,7 @@ class ModelType(Enum):
     FLOW = 6
     V_PREDICTION_CONTINUOUS = 7
     FLUX = 8
+    IMG_TO_IMG = 9
 
 
 from comfy.model_sampling import EPS, V_PREDICTION, EDM, ModelSamplingDiscrete, ModelSamplingContinuousEDM, StableCascadeSampling, ModelSamplingContinuousV
@@ -88,6 +90,8 @@ def model_sampling(model_config, model_type):
     elif model_type == ModelType.FLUX:
         c = comfy.model_sampling.CONST
         s = comfy.model_sampling.ModelSamplingFlux
+    elif model_type == ModelType.IMG_TO_IMG:
+        c = comfy.model_sampling.IMG_TO_IMG
 
     class ModelSampling(s, c):
         pass
@@ -139,6 +143,7 @@ class BaseModel(torch.nn.Module):
     def _apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
         sigma = t
         xc = self.model_sampling.calculate_input(sigma, x)
+
         if c_concat is not None:
             xc = torch.cat([xc] + [c_concat], dim=1)
 
@@ -600,6 +605,19 @@ class SDXL_instructpix2pix(IP2P, SDXL):
         else:
             self.process_ip2p_image_in = lambda image: image #diffusers ip2p
 
+class Lotus(BaseModel):
+    def extra_conds(self, **kwargs):
+        out = {}
+        cross_attn = kwargs.get("cross_attn", None)
+        out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn)
+        device = kwargs["device"]
+        task_emb = torch.tensor([1, 0]).float().to(device)
+        task_emb = torch.cat([torch.sin(task_emb), torch.cos(task_emb)]).unsqueeze(0)
+        out['y'] = comfy.conds.CONDRegular(task_emb)
+        return out
+
+    def __init__(self, model_config, model_type=ModelType.IMG_TO_IMG, device=None):
+        super().__init__(model_config, model_type, device=device)
 
 class StableCascade_C(BaseModel):
     def __init__(self, model_config, model_type=ModelType.STABLE_CASCADE, device=None):
@@ -974,7 +992,8 @@ class WAN21(BaseModel):
 
     def concat_cond(self, **kwargs):
         noise = kwargs.get("noise", None)
-        if self.diffusion_model.patch_embedding.weight.shape[1] == noise.shape[1]:
+        extra_channels = self.diffusion_model.patch_embedding.weight.shape[1] - noise.shape[1]
+        if extra_channels == 0:
             return None
 
         image = kwargs.get("concat_latent_image", None)
@@ -982,12 +1001,16 @@ class WAN21(BaseModel):
 
         if image is None:
             image = torch.zeros_like(noise)
+            shape_image = list(noise.shape)
+            shape_image[1] = extra_channels
+            image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device)
+        else:
+            image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            for i in range(0, image.shape[1], 16):
+                image[:, i: i + 16] = self.process_latent_in(image[:, i: i + 16])
+            image = utils.resize_to_batch_size(image, noise.shape[0])
 
-        image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
-        image = self.process_latent_in(image)
-        image = utils.resize_to_batch_size(image, noise.shape[0])
-
-        if not self.image_to_video:
+        if not self.image_to_video or extra_channels == image.shape[1]:
             return image
 
         mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
@@ -1013,3 +1036,18 @@ class WAN21(BaseModel):
         if clip_vision_output is not None:
             out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.penultimate_hidden_states)
         return out
+
+class Hunyuan3Dv2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan3d.model.Hunyuan3Dv2)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        guidance = kwargs.get("guidance", 5.0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+        return out

comfy/model_detection.py

@@ -154,7 +154,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["guidance_embed"] = len(guidance_keys) > 0
         return dit_config
 
-    if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys: #Flux
+    if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and '{}img_in.weight'.format(key_prefix) in state_dict_keys: #Flux
         dit_config = {}
         dit_config["image_model"] = "flux"
         dit_config["in_channels"] = 16
@@ -323,6 +323,21 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["model_type"] = "t2v"
         return dit_config
 
+    if '{}latent_in.weight'.format(key_prefix) in state_dict_keys:  # Hunyuan 3D
+        in_shape = state_dict['{}latent_in.weight'.format(key_prefix)].shape
+        dit_config = {}
+        dit_config["image_model"] = "hunyuan3d2"
+        dit_config["in_channels"] = in_shape[1]
+        dit_config["context_in_dim"] = state_dict['{}cond_in.weight'.format(key_prefix)].shape[1]
+        dit_config["hidden_size"] = in_shape[0]
+        dit_config["mlp_ratio"] = 4.0
+        dit_config["num_heads"] = 16
+        dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
+        dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
+        dit_config["qkv_bias"] = True
+        dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
@@ -667,8 +682,13 @@ def unet_config_from_diffusers_unet(state_dict, dtype=None):
             'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
             'use_temporal_attention': False, 'use_temporal_resblock': False}
 
+    LotusD = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': 4,
+            'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
+            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024, 'num_heads': 8,
+            'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
 
-    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p, SD15_diffusers_inpaint]
+    supported_models = [LotusD, SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p, SD15_diffusers_inpaint]
 
     for unet_config in supported_models:
         matches = True

comfy/model_management.py

@@ -51,6 +51,32 @@ cpu_state = CPUState.GPU
 
 total_vram = 0
 
+def get_supported_float8_types():
+    float8_types = []
+    try:
+        float8_types.append(torch.float8_e4m3fn)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e4m3fnuz)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e5m2)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e5m2fnuz)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e8m0fnu)
+    except:
+        pass
+    return float8_types
+
+FLOAT8_TYPES = get_supported_float8_types()
+
 xpu_available = False
 torch_version = ""
 try:
@@ -729,11 +755,8 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         return torch.float8_e5m2
 
     fp8_dtype = None
-    try:
-        if weight_dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
-            fp8_dtype = weight_dtype
-    except:
-        pass
+    if weight_dtype in FLOAT8_TYPES:
+        fp8_dtype = weight_dtype
 
     if fp8_dtype is not None:
         if supports_fp8_compute(device): #if fp8 compute is supported the casting is most likely not expensive

comfy/model_sampling.py

@@ -69,6 +69,15 @@ class CONST:
         sigma = sigma.view(sigma.shape[:1] + (1,) * (latent.ndim - 1))
         return latent / (1.0 - sigma)
 
+class X0(EPS):
+    def calculate_denoised(self, sigma, model_output, model_input):
+        return model_output
+
+class IMG_TO_IMG(X0):
+    def calculate_input(self, sigma, noise):
+        return noise
+
+
 class ModelSamplingDiscrete(torch.nn.Module):
     def __init__(self, model_config=None, zsnr=None):
         super().__init__()

comfy/sd.py

@@ -14,6 +14,7 @@ import comfy.ldm.genmo.vae.model
 import comfy.ldm.lightricks.vae.causal_video_autoencoder
 import comfy.ldm.cosmos.vae
 import comfy.ldm.wan.vae
+import comfy.ldm.hunyuan3d.vae
 import yaml
 import math
 
@@ -412,6 +413,17 @@ class VAE:
                 self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
                 self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype)
                 self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype)
+            elif "geo_decoder.cross_attn_decoder.ln_1.bias" in sd:
+                self.latent_dim = 1
+                ln_post = "geo_decoder.ln_post.weight" in sd
+                inner_size = sd["geo_decoder.output_proj.weight"].shape[1]
+                downsample_ratio = sd["post_kl.weight"].shape[0] // inner_size
+                mlp_expand = sd["geo_decoder.cross_attn_decoder.mlp.c_fc.weight"].shape[0] // inner_size
+                self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype)  # TODO
+                self.memory_used_decode = lambda shape, dtype: (1024 * 1024 * 1024 * 2.0) * model_management.dtype_size(dtype)  # TODO
+                ddconfig = {"embed_dim": 64, "num_freqs": 8, "include_pi": False, "heads": 16, "width": 1024, "num_decoder_layers": 16, "qkv_bias": False, "qk_norm": True, "geo_decoder_mlp_expand_ratio": mlp_expand, "geo_decoder_downsample_ratio": downsample_ratio, "geo_decoder_ln_post": ln_post}
+                self.first_stage_model = comfy.ldm.hunyuan3d.vae.ShapeVAE(**ddconfig)
+                self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
             else:
                 logging.warning("WARNING: No VAE weights detected, VAE not initalized.")
                 self.first_stage_model = None
@@ -498,7 +510,7 @@ class VAE:
         encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
         return comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.downscale_ratio, out_channels=self.latent_channels, downscale=True, index_formulas=self.downscale_index_formula, output_device=self.output_device)
 
-    def decode(self, samples_in):
+    def decode(self, samples_in, vae_options={}):
         self.throw_exception_if_invalid()
         pixel_samples = None
         try:
@@ -510,7 +522,7 @@ class VAE:
 
             for x in range(0, samples_in.shape[0], batch_number):
                 samples = samples_in[x:x+batch_number].to(self.vae_dtype).to(self.device)
-                out = self.process_output(self.first_stage_model.decode(samples).to(self.output_device).float())
+                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).float())
                 if pixel_samples is None:
                     pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
                 pixel_samples[x:x+batch_number] = out

comfy/supported_models.py

@@ -506,6 +506,22 @@ class SDXL_instructpix2pix(SDXL):
     def get_model(self, state_dict, prefix="", device=None):
         return model_base.SDXL_instructpix2pix(self, model_type=self.model_type(state_dict, prefix), device=device)
 
+class LotusD(SD20):
+    unet_config = {
+        "model_channels": 320,
+        "use_linear_in_transformer": True,
+        "use_temporal_attention": False,
+        "adm_in_channels": 4,
+        "in_channels": 4,
+    }
+
+    unet_extra_config = {
+        "num_classes": 'sequential'
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        return model_base.Lotus(self, device=device)
+
 class SD3(supported_models_base.BASE):
     unet_config = {
         "in_channels": 16,
@@ -953,12 +969,62 @@ class WAN21_I2V(WAN21_T2V):
     unet_config = {
         "image_model": "wan2.1",
         "model_type": "i2v",
+        "in_dim": 36,
     }
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.WAN21(self, image_to_video=True, device=device)
         return out
 
-models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V]
+class WAN21_FunControl2V(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "i2v",
+        "in_dim": 48,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21(self, image_to_video=False, device=device)
+        return out
+
+class Hunyuan3Dv2(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "hunyuan3d2",
+    }
+
+    unet_extra_config = {}
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 1.0,
+    }
+
+    memory_usage_factor = 3.5
+
+    clip_vision_prefix = "conditioner.main_image_encoder.model."
+    vae_key_prefix = ["vae."]
+
+    latent_format = latent_formats.Hunyuan3Dv2
+
+    def process_unet_state_dict_for_saving(self, state_dict):
+        replace_prefix = {"": "model."}
+        return utils.state_dict_prefix_replace(state_dict, replace_prefix)
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Hunyuan3Dv2(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return None
+
+class Hunyuan3Dv2mini(Hunyuan3Dv2):
+    unet_config = {
+        "image_model": "hunyuan3d2",
+        "depth": 8,
+    }
+
+    latent_format = latent_formats.Hunyuan3Dv2mini
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, Hunyuan3Dv2mini, Hunyuan3Dv2]
 
 models += [SVD_img2vid]