wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-04-15 04:52:31 +08:00
Code Issues Actions 13 Packages Projects Releases Wiki Activity

Fix CogVideoX concat_cond to handle temporal dimension and normalize channel count

Browse Source
This commit is contained in:
Talmaj Marinc 2026-04-10 19:47:18 +02:00
parent 955ce480e2
commit 8cf387b968
1 changed files with 18 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
comfy/model_base.py
View File

@ -1989,9 +1989,27 @@ class CogVideoX(BaseModel):
latent_dim = self.latent_format.latent_channels
image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
if noise.ndim == 5 and image.ndim == 5:
if image.shape[-3] < noise.shape[-3]:
image = torch.nn.functional.pad(image, (0, 0, 0, 0, 0, noise.shape[-3] - image.shape[-3]), "constant", 0)
elif image.shape[-3] > noise.shape[-3]:
image = image[:, :, :noise.shape[-3]]
for i in range(0, image.shape[1], latent_dim):
image[:, i:i + latent_dim] = self.process_latent_in(image[:, i:i + latent_dim])
image = utils.resize_to_batch_size(image, noise.shape[0])
if image.shape[1] > extra_channels:
image = image[:, :extra_channels]
elif image.shape[1] < extra_channels:
repeats = extra_channels // image.shape[1]
remainder = extra_channels % image.shape[1]
parts = [image] * repeats
if remainder > 0:
parts.append(image[:, :remainder])
image = torch.cat(parts, dim=1)
return image
def extra_conds(self, **kwargs):