Basic Hunyuan Video model support.

comfy/ldm/flux/layers.py

@@ -114,7 +114,7 @@ class Modulation(nn.Module):
 
 
 class DoubleStreamBlock(nn.Module):
-    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None):
         super().__init__()
 
         mlp_hidden_dim = int(hidden_size * mlp_ratio)
@@ -141,6 +141,7 @@ class DoubleStreamBlock(nn.Module):
             nn.GELU(approximate="tanh"),
             operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
         )
+        self.flipped_img_txt = flipped_img_txt
 
     def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None):
         img_mod1, img_mod2 = self.img_mod(vec)
@@ -160,13 +161,22 @@ class DoubleStreamBlock(nn.Module):
         txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
         txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
 
-        # run actual attention
-        attn = attention(torch.cat((txt_q, img_q), dim=2),
-                         torch.cat((txt_k, img_k), dim=2),
-                         torch.cat((txt_v, img_v), dim=2),
-                         pe=pe, mask=attn_mask)
+        if self.flipped_img_txt:
+            # run actual attention
+            attn = attention(torch.cat((img_q, txt_q), dim=2),
+                             torch.cat((img_k, txt_k), dim=2),
+                             torch.cat((img_v, txt_v), dim=2),
+                             pe=pe, mask=attn_mask)
 
-        txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
+            img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:]
+        else:
+            # run actual attention
+            attn = attention(torch.cat((txt_q, img_q), dim=2),
+                             torch.cat((txt_k, img_k), dim=2),
+                             torch.cat((txt_v, img_v), dim=2),
+                             pe=pe, mask=attn_mask)
+
+            txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
 
         # calculate the img bloks
         img = img + img_mod1.gate * self.img_attn.proj(img_attn)

comfy/ldm/hunyuan_video/model.py

@@ -0,0 +1,330 @@
+#Based on Flux code because of weird hunyuan video code license.
+
+import torch
+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
+
+
+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):
+        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)
+
+        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):
+        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)
+        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,
+    ):
+        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
+        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)
+        return x
+
+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
+        params = HunyuanVideoParams(**kwargs)
+        self.params = params
+        self.patch_size = params.patch_size
+        self.in_channels = params.in_channels
+        self.out_channels = params.out_channels
+        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=True, dtype=dtype, device=device, operations=operations)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
+        self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        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,
+                    flipped_img_txt=True,
+                    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 final_layer:
+            self.final_layer = LastLayer(self.hidden_size, self.patch_size[-1], self.out_channels, dtype=dtype, device=device, operations=operations)
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        txt_mask: Tensor,
+        timesteps: Tensor,
+        y: Tensor,
+        guidance: Tensor = None,
+        control=None,
+        transformer_options={},
+    ) -> Tensor:
+        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))
+
+        vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+
+        if self.params.guidance_embed:
+            if guidance is None:
+                raise ValueError("Didn't get guidance strength for guidance distilled model.")
+            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)
+
+        ids = torch.cat((img_ids, txt_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, img_len:] = txt_mask
+        else:
+            attn_mask = None
+
+        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["attention_mask"])
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": img, "txt": txt, "vec": vec, "pe": pe, "attention_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)
+
+            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((img, txt), 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["attention_mask"])
+                    return out
+
+                out = blocks_replace[("single_block", i)]({"img": img, "vec": vec, "pe": pe, "attention_mask": attn_mask}, {"original_block": block_wrap})
+                img = out["img"]
+            else:
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
+
+            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[:, : img_len] += add
+
+        img = img[:, : img_len]
+
+        img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
+
+        shape = initial_shape[-3:]
+        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)
+        img = img.permute(0, 4, 1, 5, 2, 6, 3, 7)
+        img = img.reshape(initial_shape)
+        return img
+
+    def forward(self, x, timestep, context, y, guidance, attention_mask=None, control=None, transformer_options={}, **kwargs):
+        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)
+        img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, guidance, control, transformer_options)
+        return out

comfy/ldm/models/autoencoder.py

@@ -162,12 +162,19 @@ class AutoencodingEngineLegacy(AutoencodingEngine):
             },
             **kwargs,
         )
-        self.quant_conv = comfy.ops.disable_weight_init.Conv2d(
+
+        if ddconfig.get("conv3d", False):
+            conv_op = comfy.ops.disable_weight_init.Conv3d
+        else:
+            conv_op = comfy.ops.disable_weight_init.Conv2d
+
+        self.quant_conv = conv_op(
             (1 + ddconfig["double_z"]) * ddconfig["z_channels"],
             (1 + ddconfig["double_z"]) * embed_dim,
             1,
         )
-        self.post_quant_conv = comfy.ops.disable_weight_init.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+
+        self.post_quant_conv = conv_op(embed_dim, ddconfig["z_channels"], 1)
         self.embed_dim = embed_dim
 
     def get_autoencoder_params(self) -> list:

comfy/ldm/modules/diffusionmodules/model.py

@@ -43,51 +43,100 @@ def Normalize(in_channels, num_groups=32):
     return ops.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
 
 
+class VideoConv3d(nn.Module):
+    def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding_mode='replicate', padding=1, **kwargs):
+        super().__init__()
+
+        self.padding_mode = padding_mode
+        if padding != 0:
+            padding = (padding, padding, padding, padding, kernel_size - 1, 0)
+        else:
+            kwargs["padding"] = padding
+
+        self.padding = padding
+        self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs)
+
+    def forward(self, x):
+        if self.padding != 0:
+            x = torch.nn.functional.pad(x, self.padding, mode=self.padding_mode)
+        return self.conv(x)
+
+def interpolate_up(x, scale_factor):
+    try:
+        return torch.nn.functional.interpolate(x, scale_factor=scale_factor, mode="nearest")
+    except: #operation not implemented for bf16
+        orig_shape = list(x.shape)
+        out_shape = orig_shape[:2]
+        for i in range(len(orig_shape) - 2):
+            out_shape.append(round(orig_shape[i + 2] * scale_factor[i]))
+        out = torch.empty(out_shape, dtype=x.dtype, layout=x.layout, device=x.device)
+        split = 8
+        l = out.shape[1] // split
+        for i in range(0, out.shape[1], l):
+            out[:,i:i+l] = torch.nn.functional.interpolate(x[:,i:i+l].to(torch.float32), scale_factor=scale_factor, mode="nearest").to(x.dtype)
+        return out
+
 class Upsample(nn.Module):
-    def __init__(self, in_channels, with_conv):
+    def __init__(self, in_channels, with_conv, conv_op=ops.Conv2d, scale_factor=2.0):
         super().__init__()
         self.with_conv = with_conv
+        self.scale_factor = scale_factor
+
         if self.with_conv:
-            self.conv = ops.Conv2d(in_channels,
+            self.conv = conv_op(in_channels,
                                         in_channels,
                                         kernel_size=3,
                                         stride=1,
                                         padding=1)
 
     def forward(self, x):
-        try:
-            x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
-        except: #operation not implemented for bf16
-            b, c, h, w = x.shape
-            out = torch.empty((b, c, h*2, w*2), dtype=x.dtype, layout=x.layout, device=x.device)
-            split = 8
-            l = out.shape[1] // split
-            for i in range(0, out.shape[1], l):
-                out[:,i:i+l] = torch.nn.functional.interpolate(x[:,i:i+l].to(torch.float32), scale_factor=2.0, mode="nearest").to(x.dtype)
-            del x
-            x = out
+        scale_factor = self.scale_factor
+        if not isinstance(scale_factor, tuple):
+            scale_factor = (scale_factor,) * (x.ndim - 2)
 
+        if x.ndim == 5 and scale_factor[0] > 1.0:
+            t = x.shape[2]
+            if t > 1:
+                a, b = x.split((1, t - 1), dim=2)
+                del x
+                b = interpolate_up(b, scale_factor)
+            else:
+                a = x
+
+            a = interpolate_up(a.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2)
+            if t > 1:
+                x = torch.cat((a, b), dim=2)
+            else:
+                x = a
+        else:
+            x = interpolate_up(x, self.scale_factor)
         if self.with_conv:
             x = self.conv(x)
         return x
 
 
 class Downsample(nn.Module):
-    def __init__(self, in_channels, with_conv):
+    def __init__(self, in_channels, with_conv, stride=2, conv_op=ops.Conv2d):
         super().__init__()
         self.with_conv = with_conv
         if self.with_conv:
             # no asymmetric padding in torch conv, must do it ourselves
-            self.conv = ops.Conv2d(in_channels,
+            self.conv = conv_op(in_channels,
                                         in_channels,
                                         kernel_size=3,
-                                        stride=2,
+                                        stride=stride,
                                         padding=0)
 
     def forward(self, x):
         if self.with_conv:
-            pad = (0,1,0,1)
-            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            if x.ndim == 4:
+                pad = (0, 1, 0, 1)
+                mode = "constant"
+                x = torch.nn.functional.pad(x, pad, mode=mode, value=0)
+            elif x.ndim == 5:
+                pad = (1, 1, 1, 1, 2, 0)
+                mode = "replicate"
+                x = torch.nn.functional.pad(x, pad, mode=mode)
             x = self.conv(x)
         else:
             x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
@@ -96,7 +145,7 @@ class Downsample(nn.Module):
 
 class ResnetBlock(nn.Module):
     def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
-                 dropout, temb_channels=512):
+                 dropout, temb_channels=512, conv_op=ops.Conv2d):
         super().__init__()
         self.in_channels = in_channels
         out_channels = in_channels if out_channels is None else out_channels
@@ -105,7 +154,7 @@ class ResnetBlock(nn.Module):
 
         self.swish = torch.nn.SiLU(inplace=True)
         self.norm1 = Normalize(in_channels)
-        self.conv1 = ops.Conv2d(in_channels,
+        self.conv1 = conv_op(in_channels,
                                      out_channels,
                                      kernel_size=3,
                                      stride=1,
@@ -115,20 +164,20 @@ class ResnetBlock(nn.Module):
                                              out_channels)
         self.norm2 = Normalize(out_channels)
         self.dropout = torch.nn.Dropout(dropout, inplace=True)
-        self.conv2 = ops.Conv2d(out_channels,
+        self.conv2 = conv_op(out_channels,
                                      out_channels,
                                      kernel_size=3,
                                      stride=1,
                                      padding=1)
         if self.in_channels != self.out_channels:
             if self.use_conv_shortcut:
-                self.conv_shortcut = ops.Conv2d(in_channels,
+                self.conv_shortcut = conv_op(in_channels,
                                                      out_channels,
                                                      kernel_size=3,
                                                      stride=1,
                                                      padding=1)
             else:
-                self.nin_shortcut = ops.Conv2d(in_channels,
+                self.nin_shortcut = conv_op(in_channels,
                                                     out_channels,
                                                     kernel_size=1,
                                                     stride=1,
@@ -194,21 +243,25 @@ def slice_attention(q, k, v):
 
 def normal_attention(q, k, v):
     # compute attention
-    b,c,h,w = q.shape
+    orig_shape = q.shape
+    b = orig_shape[0]
+    c = orig_shape[1]
 
-    q = q.reshape(b,c,h*w)
-    q = q.permute(0,2,1)   # b,hw,c
-    k = k.reshape(b,c,h*w) # b,c,hw
-    v = v.reshape(b,c,h*w)
+    q = q.reshape(b, c, -1)
+    q = q.permute(0, 2, 1)   # b,hw,c
+    k = k.reshape(b, c, -1) # b,c,hw
+    v = v.reshape(b, c, -1)
 
     r1 = slice_attention(q, k, v)
-    h_ = r1.reshape(b,c,h,w)
+    h_ = r1.reshape(orig_shape)
     del r1
     return h_
 
 def xformers_attention(q, k, v):
     # compute attention
-    B, C, H, W = q.shape
+    orig_shape = q.shape
+    B = orig_shape[0]
+    C = orig_shape[1]
     q, k, v = map(
         lambda t: t.view(B, C, -1).transpose(1, 2).contiguous(),
         (q, k, v),
@@ -216,14 +269,16 @@ def xformers_attention(q, k, v):
 
     try:
         out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None)
-        out = out.transpose(1, 2).reshape(B, C, H, W)
+        out = out.transpose(1, 2).reshape(orig_shape)
     except NotImplementedError:
-        out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(B, C, H, W)
+        out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(orig_shape)
     return out
 
 def pytorch_attention(q, k, v):
     # compute attention
-    B, C, H, W = q.shape
+    orig_shape = q.shape
+    B = orig_shape[0]
+    C = orig_shape[1]
     q, k, v = map(
         lambda t: t.view(B, 1, C, -1).transpose(2, 3).contiguous(),
         (q, k, v),
@@ -231,35 +286,35 @@ def pytorch_attention(q, k, v):
 
     try:
         out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
-        out = out.transpose(2, 3).reshape(B, C, H, W)
+        out = out.transpose(2, 3).reshape(orig_shape)
     except model_management.OOM_EXCEPTION:
         logging.warning("scaled_dot_product_attention OOMed: switched to slice attention")
-        out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(B, C, H, W)
+        out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(orig_shape)
     return out
 
 
 class AttnBlock(nn.Module):
-    def __init__(self, in_channels):
+    def __init__(self, in_channels, conv_op=ops.Conv2d):
         super().__init__()
         self.in_channels = in_channels
 
         self.norm = Normalize(in_channels)
-        self.q = ops.Conv2d(in_channels,
+        self.q = conv_op(in_channels,
                                  in_channels,
                                  kernel_size=1,
                                  stride=1,
                                  padding=0)
-        self.k = ops.Conv2d(in_channels,
+        self.k = conv_op(in_channels,
                                  in_channels,
                                  kernel_size=1,
                                  stride=1,
                                  padding=0)
-        self.v = ops.Conv2d(in_channels,
+        self.v = conv_op(in_channels,
                                  in_channels,
                                  kernel_size=1,
                                  stride=1,
                                  padding=0)
-        self.proj_out = ops.Conv2d(in_channels,
+        self.proj_out = conv_op(in_channels,
                                         in_channels,
                                         kernel_size=1,
                                         stride=1,
@@ -289,8 +344,8 @@ class AttnBlock(nn.Module):
         return x+h_
 
 
-def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
-    return AttnBlock(in_channels)
+def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None, conv_op=ops.Conv2d):
+    return AttnBlock(in_channels, conv_op=conv_op)
 
 
 class Model(nn.Module):
@@ -449,6 +504,7 @@ class Encoder(nn.Module):
     def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
                  attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
                  resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
+                 conv3d=False, time_compress=None,
                  **ignore_kwargs):
         super().__init__()
         if use_linear_attn: attn_type = "linear"
@@ -459,8 +515,15 @@ class Encoder(nn.Module):
         self.resolution = resolution
         self.in_channels = in_channels
 
+        if conv3d:
+            conv_op = VideoConv3d
+            mid_attn_conv_op = ops.Conv3d
+        else:
+            conv_op = ops.Conv2d
+            mid_attn_conv_op = ops.Conv2d
+
         # downsampling
-        self.conv_in = ops.Conv2d(in_channels,
+        self.conv_in = conv_op(in_channels,
                                        self.ch,
                                        kernel_size=3,
                                        stride=1,
@@ -479,15 +542,20 @@ class Encoder(nn.Module):
                 block.append(ResnetBlock(in_channels=block_in,
                                          out_channels=block_out,
                                          temb_channels=self.temb_ch,
-                                         dropout=dropout))
+                                         dropout=dropout,
+                                         conv_op=conv_op))
                 block_in = block_out
                 if curr_res in attn_resolutions:
-                    attn.append(make_attn(block_in, attn_type=attn_type))
+                    attn.append(make_attn(block_in, attn_type=attn_type, conv_op=conv_op))
             down = nn.Module()
             down.block = block
             down.attn = attn
             if i_level != self.num_resolutions-1:
-                down.downsample = Downsample(block_in, resamp_with_conv)
+                stride = 2
+                if time_compress is not None:
+                    if (self.num_resolutions - 1 - i_level) > math.log2(time_compress):
+                        stride = (1, 2, 2)
+                down.downsample = Downsample(block_in, resamp_with_conv, stride=stride, conv_op=conv_op)
                 curr_res = curr_res // 2
             self.down.append(down)
 
@@ -496,16 +564,18 @@ class Encoder(nn.Module):
         self.mid.block_1 = ResnetBlock(in_channels=block_in,
                                        out_channels=block_in,
                                        temb_channels=self.temb_ch,
-                                       dropout=dropout)
-        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+                                       dropout=dropout,
+                                       conv_op=conv_op)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type, conv_op=mid_attn_conv_op)
         self.mid.block_2 = ResnetBlock(in_channels=block_in,
                                        out_channels=block_in,
                                        temb_channels=self.temb_ch,
-                                       dropout=dropout)
+                                       dropout=dropout,
+                                       conv_op=conv_op)
 
         # end
         self.norm_out = Normalize(block_in)
-        self.conv_out = ops.Conv2d(block_in,
+        self.conv_out = conv_op(block_in,
                                         2*z_channels if double_z else z_channels,
                                         kernel_size=3,
                                         stride=1,
@@ -543,6 +613,8 @@ class Decoder(nn.Module):
                  conv_out_op=ops.Conv2d,
                  resnet_op=ResnetBlock,
                  attn_op=AttnBlock,
+                 conv3d=False,
+                 time_compress=None,
                 **ignorekwargs):
         super().__init__()
         self.ch = ch
@@ -554,6 +626,14 @@ class Decoder(nn.Module):
         self.give_pre_end = give_pre_end
         self.tanh_out = tanh_out
 
+        if conv3d:
+            conv_op = VideoConv3d
+            conv_out_op = VideoConv3d
+            mid_attn_conv_op = ops.Conv3d
+        else:
+            conv_op = ops.Conv2d
+            mid_attn_conv_op = ops.Conv2d
+
         # compute block_in and curr_res at lowest res
         block_in = ch*ch_mult[self.num_resolutions-1]
         curr_res = resolution // 2**(self.num_resolutions-1)
@@ -562,7 +642,7 @@ class Decoder(nn.Module):
             self.z_shape, np.prod(self.z_shape)))
 
         # z to block_in
-        self.conv_in = ops.Conv2d(z_channels,
+        self.conv_in = conv_op(z_channels,
                                        block_in,
                                        kernel_size=3,
                                        stride=1,
@@ -573,12 +653,14 @@ class Decoder(nn.Module):
         self.mid.block_1 = resnet_op(in_channels=block_in,
                                        out_channels=block_in,
                                        temb_channels=self.temb_ch,
-                                       dropout=dropout)
-        self.mid.attn_1 = attn_op(block_in)
+                                       dropout=dropout,
+                                       conv_op=conv_op)
+        self.mid.attn_1 = attn_op(block_in, conv_op=mid_attn_conv_op)
         self.mid.block_2 = resnet_op(in_channels=block_in,
                                        out_channels=block_in,
                                        temb_channels=self.temb_ch,
-                                       dropout=dropout)
+                                       dropout=dropout,
+                                       conv_op=conv_op)
 
         # upsampling
         self.up = nn.ModuleList()
@@ -590,15 +672,21 @@ class Decoder(nn.Module):
                 block.append(resnet_op(in_channels=block_in,
                                          out_channels=block_out,
                                          temb_channels=self.temb_ch,
-                                         dropout=dropout))
+                                         dropout=dropout,
+                                         conv_op=conv_op))
                 block_in = block_out
                 if curr_res in attn_resolutions:
-                    attn.append(attn_op(block_in))
+                    attn.append(attn_op(block_in, conv_op=conv_op))
             up = nn.Module()
             up.block = block
             up.attn = attn
             if i_level != 0:
-                up.upsample = Upsample(block_in, resamp_with_conv)
+                scale_factor = 2.0
+                if time_compress is not None:
+                    if i_level > math.log2(time_compress):
+                        scale_factor = (1.0, 2.0, 2.0)
+
+                up.upsample = Upsample(block_in, resamp_with_conv, conv_op=conv_op, scale_factor=scale_factor)
                 curr_res = curr_res * 2
             self.up.insert(0, up) # prepend to get consistent order
 

comfy/ldm/modules/temporal_ae.py

@@ -194,6 +194,7 @@ def make_time_attn(
     attn_kwargs=None,
     alpha: float = 0,
     merge_strategy: str = "learned",
+    conv_op=ops.Conv2d,
 ):
     return partialclass(
         AttnVideoBlock, in_channels, alpha=alpha, merge_strategy=merge_strategy