Merge branch 'master' into fix-context-window-slicing

Author: Kosinkadink

comfy/latent_formats.py

@@ -533,6 +533,79 @@ def __init__(self):
                 0.3971, 1.0600, 0.3943, 0.5537, 0.5444, 0.4089, 0.7468, 0.7744
             ]).view(1, self.latent_channels, 1, 1, 1)
 
+class HunyuanImage21(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 2
+    scale_factor = 0.75289
+
+    latent_rgb_factors = [
+        [-0.0154, -0.0397, -0.0521],
+        [ 0.0005,  0.0093,  0.0006],
+        [-0.0805, -0.0773, -0.0586],
+        [-0.0494, -0.0487, -0.0498],
+        [-0.0212, -0.0076, -0.0261],
+        [-0.0179, -0.0417, -0.0505],
+        [ 0.0158,  0.0310,  0.0239],
+        [ 0.0409,  0.0516,  0.0201],
+        [ 0.0350,  0.0553,  0.0036],
+        [-0.0447, -0.0327, -0.0479],
+        [-0.0038, -0.0221, -0.0365],
+        [-0.0423, -0.0718, -0.0654],
+        [ 0.0039,  0.0368,  0.0104],
+        [ 0.0655,  0.0217,  0.0122],
+        [ 0.0490,  0.1638,  0.2053],
+        [ 0.0932,  0.0829,  0.0650],
+        [-0.0186, -0.0209, -0.0135],
+        [-0.0080, -0.0076, -0.0148],
+        [-0.0284, -0.0201,  0.0011],
+        [-0.0642, -0.0294, -0.0777],
+        [-0.0035,  0.0076, -0.0140],
+        [ 0.0519,  0.0731,  0.0887],
+        [-0.0102,  0.0095,  0.0704],
+        [ 0.0068,  0.0218, -0.0023],
+        [-0.0726, -0.0486, -0.0519],
+        [ 0.0260,  0.0295,  0.0263],
+        [ 0.0250,  0.0333,  0.0341],
+        [ 0.0168, -0.0120, -0.0174],
+        [ 0.0226,  0.1037,  0.0114],
+        [ 0.2577,  0.1906,  0.1604],
+        [-0.0646, -0.0137, -0.0018],
+        [-0.0112,  0.0309,  0.0358],
+        [-0.0347,  0.0146, -0.0481],
+        [ 0.0234,  0.0179,  0.0201],
+        [ 0.0157,  0.0313,  0.0225],
+        [ 0.0423,  0.0675,  0.0524],
+        [-0.0031,  0.0027, -0.0255],
+        [ 0.0447,  0.0555,  0.0330],
+        [-0.0152,  0.0103,  0.0299],
+        [-0.0755, -0.0489, -0.0635],
+        [ 0.0853,  0.0788,  0.1017],
+        [-0.0272, -0.0294, -0.0471],
+        [ 0.0440,  0.0400, -0.0137],
+        [ 0.0335,  0.0317, -0.0036],
+        [-0.0344, -0.0621, -0.0984],
+        [-0.0127, -0.0630, -0.0620],
+        [-0.0648,  0.0360,  0.0924],
+        [-0.0781, -0.0801, -0.0409],
+        [ 0.0363,  0.0613,  0.0499],
+        [ 0.0238,  0.0034,  0.0041],
+        [-0.0135,  0.0258,  0.0310],
+        [ 0.0614,  0.1086,  0.0589],
+        [ 0.0428,  0.0350,  0.0205],
+        [ 0.0153,  0.0173, -0.0018],
+        [-0.0288, -0.0455, -0.0091],
+        [ 0.0344,  0.0109, -0.0157],
+        [-0.0205, -0.0247, -0.0187],
+        [ 0.0487,  0.0126,  0.0064],
+        [-0.0220, -0.0013,  0.0074],
+        [-0.0203, -0.0094, -0.0048],
+        [-0.0719,  0.0429, -0.0442],
+        [ 0.1042,  0.0497,  0.0356],
+        [-0.0659, -0.0578, -0.0280],
+        [-0.0060, -0.0322, -0.0234]]
+
+    latent_rgb_factors_bias = [0.0007, -0.0256, -0.0206]
+
 class Hunyuan3Dv2(LatentFormat):
     latent_channels = 64
     latent_dimensions = 1

comfy/ldm/hunyuan3dv2_1/hunyuandit.py

@@ -426,7 +426,7 @@ def __init__(
         text_states_dim=1024,
         qk_norm=False,
         norm_layer=nn.LayerNorm,
-        qk_norm_layer=nn.RMSNorm,
+        qk_norm_layer=True,
         qkv_bias=True,
         skip_connection=True,
         timested_modulate=False,

comfy/ldm/hunyuan_video/model.py

@@ -40,6 +40,8 @@ class HunyuanVideoParams:
     patch_size: list
     qkv_bias: bool
     guidance_embed: bool
+    byt5: bool
+    meanflow: bool
 
 
 class SelfAttentionRef(nn.Module):
@@ -161,6 +163,30 @@ def forward(
         x = self.individual_token_refiner(x, c, mask)
         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.
@@ -185,9 +211,13 @@ def __init__(self, image_model=None, final_layer=True, dtype=None, device=None,
         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.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)
-        self.vector_in = MLPEmbedder(params.vec_in_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()
         )
@@ -215,6 +245,23 @@ def __init__(self, image_model=None, final_layer=True, dtype=None, device=None,
             ]
         )
 
+        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)
 
@@ -226,7 +273,8 @@ def forward_orig(
         txt_ids: Tensor,
         txt_mask: Tensor,
         timesteps: Tensor,
-        y: Tensor,
+        y: Tensor = None,
+        txt_byt5=None,
         guidance: Tensor = None,
         guiding_frame_index=None,
         ref_latent=None,
@@ -240,6 +288,14 @@ def forward_orig(
         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:
+            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) / 2
+
         if ref_latent is not None:
             ref_latent_ids = self.img_ids(ref_latent)
             ref_latent = self.img_in(ref_latent)
@@ -250,13 +306,17 @@ def forward_orig(
 
         if guiding_frame_index is not None:
             token_replace_vec = self.time_in(timestep_embedding(guiding_frame_index, 256, time_factor=1.0))
-            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)
+            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:
-            vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+            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
 
@@ -269,6 +329,12 @@ def forward_orig(
 
         txt = self.txt_in(txt, timesteps, txt_mask)
 
+        if self.byt5_in is not None and txt_byt5 is not None:
+            txt_byt5 = self.byt5_in(txt_byt5)
+            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 = torch.cat((txt, txt_byt5), dim=1)
+            txt_ids = torch.cat((txt_ids, txt_byt5_ids), dim=1)
+
         ids = torch.cat((img_ids, txt_ids), dim=1)
         pe = self.pe_embedder(ids)
 
@@ -328,12 +394,16 @@ def block_wrap(args):
 
         img = self.final_layer(img, vec, modulation_dims=modulation_dims)  # (N, T, patch_size ** 2 * out_channels)
 
-        shape = initial_shape[-3:]
+        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)
-        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])
+        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):
@@ -348,16 +418,30 @@ def img_ids(self, x):
         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 forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, control=None, transformer_options={}, **kwargs):
+    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, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, 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, guidance, attention_mask, guiding_frame_index, ref_latent, control, transformer_options, **kwargs)
+        ).execute(x, timestep, context, y, txt_byt5, guidance, attention_mask, guiding_frame_index, ref_latent, control, transformer_options, **kwargs)
 
-    def _forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, control=None, transformer_options={}, **kwargs):
-        bs, c, t, h, w = x.shape
-        img_ids = self.img_ids(x)
-        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, guiding_frame_index, ref_latent, control=control, transformer_options=transformer_options)
+    def _forward(self, x, timestep, context, y=None, txt_byt5=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, 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, guidance, guiding_frame_index, ref_latent, control=control, transformer_options=transformer_options)
         return out