Track bundled vendor runtime sources

vendor/ComfyUI/comfy/ldm/hunyuan_video/model.py

@@ -0,0 +1,497 @@
+#Based on Flux code because of weird hunyuan video code license.
+
+import torch
+import comfy.patcher_extension
+import comfy.ldm.flux.layers
+import comfy.ldm.modules.diffusionmodules.mmdit
+from comfy.ldm.modules.attention import optimized_attention
+
+from dataclasses import dataclass
+from einops import repeat
+
+from torch import Tensor, nn
+
+from comfy.ldm.flux.layers import (
+    DoubleStreamBlock,
+    EmbedND,
+    LastLayer,
+    MLPEmbedder,
+    SingleStreamBlock,
+    timestep_embedding
+)
+
+import comfy.ldm.common_dit
+
+
+@dataclass
+class HunyuanVideoParams:
+    in_channels: int
+    out_channels: int
+    vec_in_dim: int
+    context_in_dim: int
+    hidden_size: int
+    mlp_ratio: float
+    num_heads: int
+    depth: int
+    depth_single_blocks: int
+    axes_dim: list
+    theta: int
+    patch_size: list
+    qkv_bias: bool
+    guidance_embed: bool
+    byt5: bool
+    meanflow: bool
+    use_cond_type_embedding: bool
+    vision_in_dim: int
+    meanflow_sum: bool
+
+
+class SelfAttentionRef(nn.Module):
+    def __init__(self, dim: int, qkv_bias: bool = False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+        self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+
+
+class TokenRefinerBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        heads,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.heads = heads
+        mlp_hidden_dim = hidden_size * 4
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+        self.norm1 = operations.LayerNorm(hidden_size, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device)
+        self.self_attn = SelfAttentionRef(hidden_size, True, dtype=dtype, device=device, operations=operations)
+
+        self.norm2 = operations.LayerNorm(hidden_size, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device)
+
+        self.mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+    def forward(self, x, c, mask, transformer_options={}):
+        mod1, mod2 = self.adaLN_modulation(c).chunk(2, dim=1)
+
+        norm_x = self.norm1(x)
+        qkv = self.self_attn.qkv(norm_x)
+        q, k, v = qkv.reshape(qkv.shape[0], qkv.shape[1], 3, self.heads, -1).permute(2, 0, 3, 1, 4)
+        attn = optimized_attention(q, k, v, self.heads, mask=mask, skip_reshape=True, transformer_options=transformer_options)
+
+        x = x + self.self_attn.proj(attn) * mod1.unsqueeze(1)
+        x = x + self.mlp(self.norm2(x)) * mod2.unsqueeze(1)
+        return x
+
+
+class IndividualTokenRefiner(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        heads,
+        num_blocks,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.blocks = nn.ModuleList(
+            [
+                TokenRefinerBlock(
+                    hidden_size=hidden_size,
+                    heads=heads,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations
+                )
+                for _ in range(num_blocks)
+            ]
+        )
+
+    def forward(self, x, c, mask, transformer_options={}):
+        m = None
+        if mask is not None:
+            m = mask.view(mask.shape[0], 1, 1, mask.shape[1]).repeat(1, 1, mask.shape[1], 1)
+            m = m + m.transpose(2, 3)
+
+        for block in self.blocks:
+            x = block(x, c, m, transformer_options=transformer_options)
+        return x
+
+
+
+class TokenRefiner(nn.Module):
+    def __init__(
+        self,
+        text_dim,
+        hidden_size,
+        heads,
+        num_blocks,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+
+        self.input_embedder = operations.Linear(text_dim, hidden_size, bias=True, dtype=dtype, device=device)
+        self.t_embedder = MLPEmbedder(256, hidden_size, dtype=dtype, device=device, operations=operations)
+        self.c_embedder = MLPEmbedder(text_dim, hidden_size, dtype=dtype, device=device, operations=operations)
+        self.individual_token_refiner = IndividualTokenRefiner(hidden_size, heads, num_blocks, dtype=dtype, device=device, operations=operations)
+
+    def forward(
+        self,
+        x,
+        timesteps,
+        mask,
+        transformer_options={},
+    ):
+        t = self.t_embedder(timestep_embedding(timesteps, 256, time_factor=1.0).to(x.dtype))
+        # m = mask.float().unsqueeze(-1)
+        # c = (x.float() * m).sum(dim=1) / m.sum(dim=1) #TODO: the following works when the x.shape is the same length as the tokens but might break otherwise
+        if x.dtype == torch.float16:
+            c = x.float().sum(dim=1) / x.shape[1]
+        else:
+            c = x.sum(dim=1) / x.shape[1]
+
+        c = t + self.c_embedder(c.to(x.dtype))
+        x = self.input_embedder(x)
+        x = self.individual_token_refiner(x, c, mask, transformer_options=transformer_options)
+        return x
+
+
+class ByT5Mapper(nn.Module):
+    def __init__(self, in_dim, out_dim, hidden_dim, out_dim1, use_res=False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.layernorm = operations.LayerNorm(in_dim, dtype=dtype, device=device)
+        self.fc1 = operations.Linear(in_dim, hidden_dim, dtype=dtype, device=device)
+        self.fc2 = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device)
+        self.fc3 = operations.Linear(out_dim, out_dim1, dtype=dtype, device=device)
+        self.use_res = use_res
+        self.act_fn = nn.GELU()
+
+    def forward(self, x):
+        if self.use_res:
+            res = x
+        x = self.layernorm(x)
+        x = self.fc1(x)
+        x = self.act_fn(x)
+        x = self.fc2(x)
+        x2 = self.act_fn(x)
+        x2 = self.fc3(x2)
+        if self.use_res:
+            x2 = x2 + res
+        return x2
+
+class HunyuanVideo(nn.Module):
+    """
+    Transformer model for flow matching on sequences.
+    """
+
+    def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
+        super().__init__()
+        self.dtype = dtype
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        params = HunyuanVideoParams(**kwargs)
+        self.params = params
+        self.patch_size = params.patch_size
+        self.in_channels = params.in_channels
+        self.out_channels = params.out_channels
+        self.use_cond_type_embedding = params.use_cond_type_embedding
+        self.vision_in_dim = params.vision_in_dim
+        if params.hidden_size % params.num_heads != 0:
+            raise ValueError(
+                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
+            )
+        pe_dim = params.hidden_size // params.num_heads
+        if sum(params.axes_dim) != pe_dim:
+            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = params.hidden_size
+        self.num_heads = params.num_heads
+        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
+
+        self.img_in = comfy.ldm.modules.diffusionmodules.mmdit.PatchEmbed(None, self.patch_size, self.in_channels, self.hidden_size, conv3d=len(self.patch_size) == 3, dtype=dtype, device=device, operations=operations)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
+        if params.vec_in_dim is not None:
+            self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        else:
+            self.vector_in = None
+
+        self.guidance_in = (
+            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
+        )
+
+        self.txt_in = TokenRefiner(params.context_in_dim, self.hidden_size, self.num_heads, 2, dtype=dtype, device=device, operations=operations)
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(params.depth)
+            ]
+        )
+
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                for _ in range(params.depth_single_blocks)
+            ]
+        )
+
+        if params.byt5:
+            self.byt5_in = ByT5Mapper(
+                in_dim=1472,
+                out_dim=2048,
+                hidden_dim=2048,
+                out_dim1=self.hidden_size,
+                use_res=False,
+                dtype=dtype, device=device, operations=operations
+            )
+        else:
+            self.byt5_in = None
+
+        if params.meanflow:
+            self.time_r_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
+        else:
+            self.time_r_in = None
+
+        if final_layer:
+            self.final_layer = LastLayer(self.hidden_size, self.patch_size[-1], self.out_channels, dtype=dtype, device=device, operations=operations)
+
+        # HunyuanVideo 1.5 specific modules
+        if self.vision_in_dim is not None:
+            from comfy.ldm.wan.model import MLPProj
+            self.vision_in = MLPProj(in_dim=self.vision_in_dim, out_dim=self.hidden_size, operation_settings=operation_settings)
+        else:
+            self.vision_in = None
+        if self.use_cond_type_embedding:
+            # 0: text_encoder feature 1: byt5 feature 2: vision_encoder feature
+            self.cond_type_embedding = nn.Embedding(3, self.hidden_size)
+        else:
+            self.cond_type_embedding = None
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        txt_mask: Tensor,
+        timesteps: Tensor,
+        y: Tensor = None,
+        txt_byt5=None,
+        clip_fea=None,
+        guidance: Tensor = None,
+        guiding_frame_index=None,
+        ref_latent=None,
+        disable_time_r=False,
+        control=None,
+        transformer_options={},
+    ) -> Tensor:
+        transformer_options = transformer_options.copy()
+        patches_replace = transformer_options.get("patches_replace", {})
+
+        initial_shape = list(img.shape)
+        # running on sequences img
+        img = self.img_in(img)
+        vec = self.time_in(timestep_embedding(timesteps, 256, time_factor=1.0).to(img.dtype))
+
+        if (self.time_r_in is not None) and (not disable_time_r):
+            w = torch.where(transformer_options['sigmas'][0] == transformer_options['sample_sigmas'])[0]  # This most likely could be improved
+            if len(w) > 0:
+                timesteps_r = transformer_options['sample_sigmas'][w[0] + 1]
+                timesteps_r = timesteps_r.unsqueeze(0).to(device=timesteps.device, dtype=timesteps.dtype)
+                vec_r = self.time_r_in(timestep_embedding(timesteps_r, 256, time_factor=1000.0).to(img.dtype))
+                vec = (vec + vec_r) if self.params.meanflow_sum else (vec + vec_r) / 2
+
+        if ref_latent is not None:
+            ref_latent_ids = self.img_ids(ref_latent)
+            ref_latent = self.img_in(ref_latent)
+            img = torch.cat([ref_latent, img], dim=-2)
+            ref_latent_ids[..., 0] = -1
+            ref_latent_ids[..., 2] += (initial_shape[-1] // self.patch_size[-1])
+            img_ids = torch.cat([ref_latent_ids, img_ids], dim=-2)
+
+        if guiding_frame_index is not None:
+            token_replace_vec = self.time_in(timestep_embedding(guiding_frame_index, 256, time_factor=1.0))
+            if self.vector_in is not None:
+                vec_ = self.vector_in(y[:, :self.params.vec_in_dim])
+                vec = torch.cat([(vec_ + token_replace_vec).unsqueeze(1), (vec_ + vec).unsqueeze(1)], dim=1)
+            else:
+                vec = torch.cat([(token_replace_vec).unsqueeze(1), (vec).unsqueeze(1)], dim=1)
+            frame_tokens = (initial_shape[-1] // self.patch_size[-1]) * (initial_shape[-2] // self.patch_size[-2])
+            modulation_dims = [(0, frame_tokens, 0), (frame_tokens, None, 1)]
+            modulation_dims_txt = [(0, None, 1)]
+        else:
+            if self.vector_in is not None:
+                vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+            modulation_dims = None
+            modulation_dims_txt = None
+
+        if self.params.guidance_embed:
+            if guidance is not None:
+                vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
+
+        if txt_mask is not None and not torch.is_floating_point(txt_mask):
+            txt_mask = (txt_mask - 1).to(img.dtype) * torch.finfo(img.dtype).max
+
+        txt = self.txt_in(txt, timesteps, txt_mask, transformer_options=transformer_options)
+
+        if self.cond_type_embedding is not None:
+            self.cond_type_embedding.to(txt.device)
+            cond_emb = self.cond_type_embedding(torch.zeros_like(txt[:, :, 0], device=txt.device, dtype=torch.long))
+            txt = txt + cond_emb.to(txt.dtype)
+
+        if self.byt5_in is not None and txt_byt5 is not None:
+            txt_byt5 = self.byt5_in(txt_byt5)
+            if self.cond_type_embedding is not None:
+                cond_emb = self.cond_type_embedding(torch.ones_like(txt_byt5[:, :, 0], device=txt_byt5.device, dtype=torch.long))
+                txt_byt5 = txt_byt5 + cond_emb.to(txt_byt5.dtype)
+                txt = torch.cat((txt_byt5, txt), dim=1) # byt5 first for HunyuanVideo1.5
+            else:
+                txt = torch.cat((txt, txt_byt5), dim=1)
+            txt_byt5_ids = torch.zeros((txt_ids.shape[0], txt_byt5.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype)
+            txt_ids = torch.cat((txt_ids, txt_byt5_ids), dim=1)
+
+        if clip_fea is not None:
+            txt_vision_states = self.vision_in(clip_fea)
+            if self.cond_type_embedding is not None:
+                cond_emb = self.cond_type_embedding(2 * torch.ones_like(txt_vision_states[:, :, 0], dtype=torch.long, device=txt_vision_states.device))
+                txt_vision_states = txt_vision_states + cond_emb
+            txt = torch.cat((txt_vision_states.to(txt.dtype), txt), dim=1)
+            extra_txt_ids = torch.zeros((txt_ids.shape[0], txt_vision_states.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype)
+            txt_ids = torch.cat((txt_ids, extra_txt_ids), dim=1)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        img_len = img.shape[1]
+        if txt_mask is not None:
+            attn_mask_len = img_len + txt.shape[1]
+            attn_mask = torch.zeros((1, 1, attn_mask_len), dtype=img.dtype, device=img.device)
+            attn_mask[:, 0, :txt.shape[1]] = txt_mask
+        else:
+            attn_mask = None
+
+        blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
+        for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = i
+            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["attention_mask"], modulation_dims_img=args["modulation_dims_img"], modulation_dims_txt=args["modulation_dims_txt"], transformer_options=args["transformer_options"])
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": img, "txt": txt, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims_img': modulation_dims, 'modulation_dims_txt': modulation_dims_txt, 'transformer_options': transformer_options}, {"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, modulation_dims_img=modulation_dims, modulation_dims_txt=modulation_dims_txt, transformer_options=transformer_options)
+
+            if control is not None: # Controlnet
+                control_i = control.get("input")
+                if i < len(control_i):
+                    add = control_i[i]
+                    if add is not None:
+                        img += add
+
+        img = torch.cat((txt, img), 1)
+
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
+        transformer_options["img_slice"] = [txt.shape[1], img.shape[1]]
+        for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = i
+            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["attention_mask"], modulation_dims=args["modulation_dims"], transformer_options=args["transformer_options"])
+                    return out
+
+                out = blocks_replace[("single_block", i)]({"img": img, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims': modulation_dims, 'transformer_options': transformer_options}, {"original_block": block_wrap})
+                img = out["img"]
+            else:
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims=modulation_dims, transformer_options=transformer_options)
+
+            if control is not None: # Controlnet
+                control_o = control.get("output")
+                if i < len(control_o):
+                    add = control_o[i]
+                    if add is not None:
+                        img[:, txt.shape[1]: img_len + txt.shape[1]] += add
+
+        img = img[:, txt.shape[1]: img_len + txt.shape[1]]
+        if ref_latent is not None:
+            img = img[:, ref_latent.shape[1]:]
+
+        img = self.final_layer(img, vec, modulation_dims=modulation_dims)  # (N, T, patch_size ** 2 * out_channels)
+
+        shape = initial_shape[-len(self.patch_size):]
+        for i in range(len(shape)):
+            shape[i] = shape[i] // self.patch_size[i]
+        img = img.reshape([img.shape[0]] + shape + [self.out_channels] + self.patch_size)
+        if img.ndim == 8:
+            img = img.permute(0, 4, 1, 5, 2, 6, 3, 7)
+            img = img.reshape(initial_shape[0], self.out_channels, initial_shape[2], initial_shape[3], initial_shape[4])
+        else:
+            img = img.permute(0, 3, 1, 4, 2, 5)
+            img = img.reshape(initial_shape[0], self.out_channels, initial_shape[2], initial_shape[3])
+        return img
+
+    def img_ids(self, x):
+        bs, c, t, h, w = x.shape
+        patch_size = self.patch_size
+        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
+        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
+        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
+        return repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+
+    def img_ids_2d(self, x):
+        bs, c, h, w = x.shape
+        patch_size = self.patch_size
+        h_len = ((h + (patch_size[0] // 2)) // patch_size[0])
+        w_len = ((w + (patch_size[1] // 2)) // patch_size[1])
+        img_ids = torch.zeros((h_len, w_len, 2), device=x.device, dtype=x.dtype)
+        img_ids[:, :, 0] = img_ids[:, :, 0] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        return repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+    def forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, y, txt_byt5, clip_fea, guidance, attention_mask, guiding_frame_index, ref_latent, disable_time_r, control, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
+        bs = x.shape[0]
+        if len(self.patch_size) == 3:
+            img_ids = self.img_ids(x)
+            txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        else:
+            img_ids = self.img_ids_2d(x)
+            txt_ids = torch.zeros((bs, context.shape[1], 2), device=x.device, dtype=x.dtype)
+        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, txt_byt5, clip_fea, guidance, guiding_frame_index, ref_latent, disable_time_r=disable_time_r, control=control, transformer_options=transformer_options)
+        return out

vendor/ComfyUI/comfy/ldm/hunyuan_video/upsampler.py

@@ -0,0 +1,122 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d
+from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm
+import comfy.model_management
+import comfy.model_patcher
+
+class SRResidualCausalBlock3D(nn.Module):
+    def __init__(self, channels: int):
+        super().__init__()
+        self.block = nn.Sequential(
+            VideoConv3d(channels, channels, kernel_size=3),
+            nn.SiLU(inplace=True),
+            VideoConv3d(channels, channels, kernel_size=3),
+            nn.SiLU(inplace=True),
+            VideoConv3d(channels, channels, kernel_size=3),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.block(x)
+
+class SRModel3DV2(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        hidden_channels: int = 64,
+        num_blocks: int = 6,
+        global_residual: bool = False,
+    ):
+        super().__init__()
+        self.in_conv = VideoConv3d(in_channels, hidden_channels, kernel_size=3)
+        self.blocks = nn.ModuleList([SRResidualCausalBlock3D(hidden_channels) for _ in range(num_blocks)])
+        self.out_conv = VideoConv3d(hidden_channels, out_channels, kernel_size=3)
+        self.global_residual = bool(global_residual)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        residual = x
+        y = self.in_conv(x)
+        for blk in self.blocks:
+            y = blk(y)
+        y = self.out_conv(y)
+        if self.global_residual and (y.shape == residual.shape):
+            y = y + residual
+        return y
+
+
+class Upsampler(nn.Module):
+    def __init__(
+        self,
+        z_channels: int,
+        out_channels: int,
+        block_out_channels: tuple[int, ...],
+        num_res_blocks: int = 2,
+    ):
+        super().__init__()
+        self.num_res_blocks = num_res_blocks
+        self.block_out_channels = block_out_channels
+        self.z_channels = z_channels
+
+        ch = block_out_channels[0]
+        self.conv_in = VideoConv3d(z_channels, ch, kernel_size=3)
+
+        self.up = nn.ModuleList()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                    out_channels=tgt,
+                                                    temb_channels=0,
+                                                    conv_shortcut=False,
+                                                    conv_op=VideoConv3d, norm_op=RMS_norm)
+                                        for j in range(num_res_blocks + 1)])
+            ch = tgt
+            self.up.append(stage)
+
+        self.norm_out = RMS_norm(ch)
+        self.conv_out = VideoConv3d(ch, out_channels, kernel_size=3)
+
+    def forward(self, z):
+        """
+        Args:
+            z: (B, C, T, H, W)
+            target_shape: (H, W)
+        """
+        # z to block_in
+        repeats = self.block_out_channels[0] // (self.z_channels)
+        x = self.conv_in(z) + z.repeat_interleave(repeats=repeats, dim=1)
+
+        # upsampling
+        for stage in self.up:
+            for blk in stage.block:
+                x = blk(x)
+
+        out = self.conv_out(F.silu(self.norm_out(x)))
+        return out
+
+UPSAMPLERS = {
+    "720p": SRModel3DV2,
+    "1080p": Upsampler,
+}
+
+class HunyuanVideo15SRModel():
+    def __init__(self, model_type, config):
+        self.load_device = comfy.model_management.vae_device()
+        offload_device = comfy.model_management.vae_offload_device()
+        self.dtype = comfy.model_management.vae_dtype(self.load_device)
+        self.model_class = UPSAMPLERS.get(model_type)
+        self.model = self.model_class(**config).eval()
+
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+
+    def load_sd(self, sd):
+        return self.model.load_state_dict(sd, strict=True, assign=self.patcher.is_dynamic())
+
+    def get_sd(self):
+        return self.model.state_dict()
+
+    def resample_latent(self, latent):
+        comfy.model_management.load_model_gpu(self.patcher)
+        return self.model(latent.to(self.load_device))

vendor/ComfyUI/comfy/ldm/hunyuan_video/vae.py

@@ -0,0 +1,136 @@
+import torch.nn as nn
+import torch.nn.functional as F
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+class PixelShuffle2D(nn.Module):
+    def __init__(self, in_dim, out_dim, op=ops.Conv2d):
+        super().__init__()
+        self.conv = op(in_dim, out_dim >> 2, 3, 1, 1)
+        self.ratio = (in_dim << 2) // out_dim
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        h2, w2 = h >> 1, w >> 1
+        y = self.conv(x).view(b, -1, h2, 2, w2, 2).permute(0, 3, 5, 1, 2, 4).reshape(b, -1, h2, w2)
+        r = x.view(b, c, h2, 2, w2, 2).permute(0, 3, 5, 1, 2, 4).reshape(b, c << 2, h2, w2)
+        return y + r.view(b, y.shape[1], self.ratio, h2, w2).mean(2)
+
+
+class PixelUnshuffle2D(nn.Module):
+    def __init__(self, in_dim, out_dim, op=ops.Conv2d):
+        super().__init__()
+        self.conv = op(in_dim, out_dim << 2, 3, 1, 1)
+        self.scale = (out_dim << 2) // in_dim
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        h2, w2 = h << 1, w << 1
+        y = self.conv(x).view(b, 2, 2, -1, h, w).permute(0, 3, 4, 1, 5, 2).reshape(b, -1, h2, w2)
+        r = x.repeat_interleave(self.scale, 1).view(b, 2, 2, -1, h, w).permute(0, 3, 4, 1, 5, 2).reshape(b, -1, h2, w2)
+        return y + r
+
+
+class Encoder(nn.Module):
+    def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
+                 ffactor_spatial, downsample_match_channel=True, **_):
+        super().__init__()
+        self.z_channels = z_channels
+        self.block_out_channels = block_out_channels
+        self.num_res_blocks = num_res_blocks
+        self.conv_in = ops.Conv2d(in_channels, block_out_channels[0], 3, 1, 1)
+
+        self.down = nn.ModuleList()
+        ch = block_out_channels[0]
+        depth = (ffactor_spatial >> 1).bit_length()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=ops.Conv2d)
+                                        for j in range(num_res_blocks)])
+            ch = tgt
+            if i < depth:
+                nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and downsample_match_channel else ch
+                stage.downsample = PixelShuffle2D(ch, nxt, ops.Conv2d)
+                ch = nxt
+            self.down.append(stage)
+
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=ops.Conv2d)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv2d)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=ops.Conv2d)
+
+        self.norm_out = ops.GroupNorm(32, ch, 1e-6, True)
+        self.conv_out = ops.Conv2d(ch, z_channels << 1, 3, 1, 1)
+
+    def forward(self, x):
+        x = self.conv_in(x)
+
+        for stage in self.down:
+            for blk in stage.block:
+                x = blk(x)
+            if hasattr(stage, 'downsample'):
+                x = stage.downsample(x)
+
+        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
+
+        b, c, h, w = x.shape
+        grp = c // (self.z_channels << 1)
+        skip = x.view(b, c // grp, grp, h, w).mean(2)
+
+        return self.conv_out(F.silu(self.norm_out(x))) + skip
+
+
+class Decoder(nn.Module):
+    def __init__(self, z_channels, out_channels, block_out_channels, num_res_blocks,
+                 ffactor_spatial, upsample_match_channel=True, **_):
+        super().__init__()
+        block_out_channels = block_out_channels[::-1]
+        self.z_channels = z_channels
+        self.block_out_channels = block_out_channels
+        self.num_res_blocks = num_res_blocks
+
+        ch = block_out_channels[0]
+        self.conv_in = ops.Conv2d(z_channels, ch, 3, 1, 1)
+
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=ops.Conv2d)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv2d)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=ops.Conv2d)
+
+        self.up = nn.ModuleList()
+        depth = (ffactor_spatial >> 1).bit_length()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=ops.Conv2d)
+                                        for j in range(num_res_blocks + 1)])
+            ch = tgt
+            if i < depth:
+                nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and upsample_match_channel else ch
+                stage.upsample = PixelUnshuffle2D(ch, nxt, ops.Conv2d)
+                ch = nxt
+            self.up.append(stage)
+
+        self.norm_out = ops.GroupNorm(32, ch, 1e-6, True)
+        self.conv_out = ops.Conv2d(ch, out_channels, 3, 1, 1)
+
+    def forward(self, z):
+        x = self.conv_in(z) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1)
+        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
+
+        for stage in self.up:
+            for blk in stage.block:
+                x = blk(x)
+            if hasattr(stage, 'upsample'):
+                x = stage.upsample(x)
+
+        return self.conv_out(F.silu(self.norm_out(x)))

vendor/ComfyUI/comfy/ldm/hunyuan_video/vae_refiner.py

@@ -0,0 +1,313 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, CarriedConv3d, Normalize, conv_carry_causal_3d, torch_cat_if_needed
+import comfy.ops
+import comfy.ldm.models.autoencoder
+import comfy.model_management
+ops = comfy.ops.disable_weight_init
+
+
+class RMS_norm(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        shape = (dim, 1, 1, 1)
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.empty(shape))
+
+    def forward(self, x):
+        return F.normalize(x, dim=1) * self.scale * comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device)
+
+class DnSmpl(nn.Module):
+    def __init__(self, ic, oc, tds, refiner_vae, op):
+        super().__init__()
+        fct = 2 * 2 * 2 if tds else 1 * 2 * 2
+        assert oc % fct == 0
+        self.conv = op(ic, oc // fct, kernel_size=3, stride=1, padding=1)
+        self.refiner_vae = refiner_vae
+
+        self.tds = tds
+        self.gs = fct * ic // oc
+
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
+        r1 = 2 if self.tds else 1
+        h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
+
+        if self.tds and self.refiner_vae and conv_carry_in is None:
+
+            hf = h[:, :, :1, :, :]
+            b, c, f, ht, wd = hf.shape
+            hf = hf.reshape(b, c, f, ht // 2, 2, wd // 2, 2)
+            hf = hf.permute(0, 4, 6, 1, 2, 3, 5)
+            hf = hf.reshape(b, 2 * 2 * c, f, ht // 2, wd // 2)
+            hf = torch.cat([hf, hf], dim=1)
+
+            h = h[:, :, 1:, :, :]
+
+            xf = x[:, :, :1, :, :]
+            b, ci, f, ht, wd = xf.shape
+            xf = xf.reshape(b, ci, f, ht // 2, 2, wd // 2, 2)
+            xf = xf.permute(0, 4, 6, 1, 2, 3, 5)
+            xf = xf.reshape(b, 2 * 2 * ci, f, ht // 2, wd // 2)
+            B, C, T, H, W = xf.shape
+            xf = xf.view(B, hf.shape[1], self.gs // 2, T, H, W).mean(dim=2)
+
+            x = x[:, :, 1:, :, :]
+
+        if h.shape[2] == 0:
+            return hf + xf
+
+        b, c, frms, ht, wd = h.shape
+        nf = frms // r1
+        h = h.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2)
+        h = h.permute(0, 3, 5, 7, 1, 2, 4, 6)
+        h = h.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2)
+
+        b, ci, frms, ht, wd = x.shape
+        nf = frms // r1
+        x = x.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2)
+        x = x.permute(0, 3, 5, 7, 1, 2, 4, 6)
+        x = x.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2)
+        B, C, T, H, W = x.shape
+        x = x.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2)
+
+        if self.tds and self.refiner_vae and conv_carry_in is None:
+            h = torch.cat([hf, h], dim=2)
+            x = torch.cat([xf, x], dim=2)
+
+        return h + x
+
+
+class UpSmpl(nn.Module):
+    def __init__(self, ic, oc, tus, refiner_vae, op):
+        super().__init__()
+        fct = 2 * 2 * 2 if tus else 1 * 2 * 2
+        self.conv = op(ic, oc * fct, kernel_size=3, stride=1, padding=1)
+        self.refiner_vae = refiner_vae
+
+        self.tus = tus
+        self.rp = fct * oc // ic
+
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
+        r1 = 2 if self.tus else 1
+        h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
+
+        if self.tus and self.refiner_vae and conv_carry_in is None:
+            hf = h[:, :, :1, :, :]
+            b, c, f, ht, wd = hf.shape
+            nc = c // (2 * 2)
+            hf = hf.reshape(b, 2, 2, nc, f, ht, wd)
+            hf = hf.permute(0, 3, 4, 5, 1, 6, 2)
+            hf = hf.reshape(b, nc, f, ht * 2, wd * 2)
+            hf = hf[:, : hf.shape[1] // 2]
+
+            h = h[:, :, 1:, :, :]
+
+            xf = x[:, :, :1, :, :]
+            b, ci, f, ht, wd = xf.shape
+            xf = xf.repeat_interleave(repeats=self.rp // 2, dim=1)
+            b, c, f, ht, wd = xf.shape
+            nc = c // (2 * 2)
+            xf = xf.reshape(b, 2, 2, nc, f, ht, wd)
+            xf = xf.permute(0, 3, 4, 5, 1, 6, 2)
+            xf = xf.reshape(b, nc, f, ht * 2, wd * 2)
+
+            x = x[:, :, 1:, :, :]
+
+        b, c, frms, ht, wd = h.shape
+        nc = c // (r1 * 2 * 2)
+        h = h.reshape(b, r1, 2, 2, nc, frms, ht, wd)
+        h = h.permute(0, 4, 5, 1, 6, 2, 7, 3)
+        h = h.reshape(b, nc, frms * r1, ht * 2, wd * 2)
+
+        x = x.repeat_interleave(repeats=self.rp, dim=1)
+        b, c, frms, ht, wd = x.shape
+        nc = c // (r1 * 2 * 2)
+        x = x.reshape(b, r1, 2, 2, nc, frms, ht, wd)
+        x = x.permute(0, 4, 5, 1, 6, 2, 7, 3)
+        x = x.reshape(b, nc, frms * r1, ht * 2, wd * 2)
+
+        if self.tus and self.refiner_vae and conv_carry_in is None:
+            h = torch.cat([hf, h], dim=2)
+            x = torch.cat([xf, x], dim=2)
+
+        return h + x
+
+class Encoder(nn.Module):
+    def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
+                 ffactor_spatial, ffactor_temporal, downsample_match_channel=True, refiner_vae=True, **_):
+        super().__init__()
+        self.z_channels = z_channels
+        self.block_out_channels = block_out_channels
+        self.num_res_blocks = num_res_blocks
+        self.ffactor_temporal = ffactor_temporal
+
+        self.refiner_vae = refiner_vae
+        if self.refiner_vae:
+            conv_op = CarriedConv3d
+            norm_op = RMS_norm
+        else:
+            conv_op = ops.Conv3d
+            norm_op = Normalize
+
+        self.conv_in = conv_op(in_channels, block_out_channels[0], 3, 1, 1)
+
+        self.down = nn.ModuleList()
+        ch = block_out_channels[0]
+        depth = (ffactor_spatial >> 1).bit_length()
+        depth_temporal = ((ffactor_spatial // self.ffactor_temporal) >> 1).bit_length()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=conv_op, norm_op=norm_op)
+                                        for j in range(num_res_blocks)])
+            ch = tgt
+            if i < depth:
+                nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and downsample_match_channel else ch
+                stage.downsample = DnSmpl(ch, nxt, tds=i >= depth_temporal, refiner_vae=self.refiner_vae, op=conv_op)
+                ch = nxt
+            self.down.append(stage)
+
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+
+        self.norm_out = norm_op(ch)
+        self.conv_out = conv_op(ch, z_channels << 1, 3, 1, 1)
+
+        self.regul = comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer()
+
+    def forward(self, x):
+        if not self.refiner_vae and x.shape[2] == 1:
+            x = x.expand(-1, -1, self.ffactor_temporal, -1, -1)
+
+        if self.refiner_vae:
+            xl = [x[:, :, :1, :, :]]
+            if x.shape[2] > self.ffactor_temporal:
+                xl += torch.split(x[:, :, 1: 1 + ((x.shape[2] - 1) // self.ffactor_temporal) * self.ffactor_temporal, :, :], self.ffactor_temporal * 2, dim=2)
+            x = xl
+        else:
+            x = [x]
+        out = []
+
+        conv_carry_in = None
+
+        for i, x1 in enumerate(x):
+            conv_carry_out = []
+            if i == len(x) - 1:
+                conv_carry_out = None
+
+            x1 = [ x1 ]
+            x1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out)
+
+            for stage in self.down:
+                for blk in stage.block:
+                    x1 = blk(x1, None, conv_carry_in, conv_carry_out)
+                if hasattr(stage, 'downsample'):
+                    x1 = stage.downsample(x1, conv_carry_in, conv_carry_out)
+
+            out.append(x1)
+            conv_carry_in = conv_carry_out
+
+        out = torch_cat_if_needed(out, dim=2)
+
+        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(out)))
+        del out
+
+        b, c, t, h, w = x.shape
+        grp = c // (self.z_channels << 1)
+        skip = x.view(b, c // grp, grp, t, h, w).mean(2)
+
+        out = conv_carry_causal_3d([F.silu(self.norm_out(x))], self.conv_out) + skip
+
+        if self.refiner_vae:
+            out = self.regul(out)[0]
+
+        return out
+
+class Decoder(nn.Module):
+    def __init__(self, z_channels, out_channels, block_out_channels, num_res_blocks,
+                 ffactor_spatial, ffactor_temporal, upsample_match_channel=True, refiner_vae=True, **_):
+        super().__init__()
+        block_out_channels = block_out_channels[::-1]
+        self.z_channels = z_channels
+        self.block_out_channels = block_out_channels
+        self.num_res_blocks = num_res_blocks
+
+        self.refiner_vae = refiner_vae
+        if self.refiner_vae:
+            conv_op = CarriedConv3d
+            norm_op = RMS_norm
+        else:
+            conv_op = ops.Conv3d
+            norm_op = Normalize
+
+        ch = block_out_channels[0]
+        self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1)
+
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch,  conv_op=conv_op, norm_op=norm_op)
+
+        self.up = nn.ModuleList()
+        depth = (ffactor_spatial >> 1).bit_length()
+        depth_temporal = (ffactor_temporal >> 1).bit_length()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=conv_op, norm_op=norm_op)
+                                        for j in range(num_res_blocks + 1)])
+            ch = tgt
+            if i < depth:
+                nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and upsample_match_channel else ch
+                stage.upsample = UpSmpl(ch, nxt, tus=i < depth_temporal, refiner_vae=self.refiner_vae, op=conv_op)
+                ch = nxt
+            self.up.append(stage)
+
+        self.norm_out = norm_op(ch)
+        self.conv_out = conv_op(ch, out_channels, 3, stride=1, padding=1)
+
+    def forward(self, z):
+        x = conv_carry_causal_3d([z], self.conv_in) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1)
+        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
+
+        if self.refiner_vae:
+            x = torch.split(x, 2, dim=2)
+        else:
+            x = [ x ]
+        out = []
+
+        conv_carry_in = None
+
+        for i, x1 in enumerate(x):
+            conv_carry_out = []
+            if i == len(x) - 1:
+                conv_carry_out = None
+            for stage in self.up:
+                for blk in stage.block:
+                    x1 = blk(x1, None, conv_carry_in, conv_carry_out)
+                if hasattr(stage, 'upsample'):
+                    x1 = stage.upsample(x1, conv_carry_in, conv_carry_out)
+
+            x1 = [ F.silu(self.norm_out(x1)) ]
+            x1 = conv_carry_causal_3d(x1, self.conv_out, conv_carry_in, conv_carry_out)
+            out.append(x1)
+            conv_carry_in = conv_carry_out
+        del x
+
+        out = torch_cat_if_needed(out, dim=2)
+
+        if not self.refiner_vae:
+            if z.shape[-3] == 1:
+                out = out[:, :, -1:]
+
+        return out
+