Skip to content

Fix batch size above 1 giving bad output in chroma radiance. #10394

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Oct 19, 2025
Merged

Fix batch size above 1 giving bad output in chroma radiance. #10394

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 7 additions & 16 deletions comfy/ldm/chroma_radiance/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -189,15 +189,15 @@ def forward_nerf(
nerf_pixels = nn.functional.unfold(img_orig, kernel_size=patch_size, stride=patch_size)
nerf_pixels = nerf_pixels.transpose(1, 2) # -> [B, NumPatches, C * P * P]

# Reshape for per-patch processing
nerf_hidden = img_out.reshape(B * num_patches, params.hidden_size)
nerf_pixels = nerf_pixels.reshape(B * num_patches, C, patch_size**2).transpose(1, 2)

if params.nerf_tile_size > 0 and num_patches > params.nerf_tile_size:
# Enable tiling if nerf_tile_size isn't 0 and we actually have more patches than
# the tile size.
img_dct = self.forward_tiled_nerf(img_out, nerf_pixels, B, C, num_patches, patch_size, params)
img_dct = self.forward_tiled_nerf(nerf_hidden, nerf_pixels, B, C, num_patches, patch_size, params)
else:
# Reshape for per-patch processing
nerf_hidden = img_out.reshape(B * num_patches, params.hidden_size)
nerf_pixels = nerf_pixels.reshape(B * num_patches, C, patch_size**2).transpose(1, 2)

# Get DCT-encoded pixel embeddings [pixel-dct]
img_dct = self.nerf_image_embedder(nerf_pixels)

Expand Down Expand Up @@ -240,17 +240,8 @@ def forward_tiled_nerf(
end = min(i + tile_size, num_patches)

# Slice the current tile from the input tensors
nerf_hidden_tile = nerf_hidden[:, i:end, :]
nerf_pixels_tile = nerf_pixels[:, i:end, :]

# Get the actual number of patches in this tile (can be smaller for the last tile)
num_patches_tile = nerf_hidden_tile.shape[1]

# Reshape the tile for per-patch processing
# [B, NumPatches_tile, D] -> [B * NumPatches_tile, D]
nerf_hidden_tile = nerf_hidden_tile.reshape(batch * num_patches_tile, params.hidden_size)
# [B, NumPatches_tile, C*P*P] -> [B*NumPatches_tile, C, P*P] -> [B*NumPatches_tile, P*P, C]
nerf_pixels_tile = nerf_pixels_tile.reshape(batch * num_patches_tile, channels, patch_size**2).transpose(1, 2)
nerf_hidden_tile = nerf_hidden[i * batch:end * batch]
nerf_pixels_tile = nerf_pixels[i * batch:end * batch]

# get DCT-encoded pixel embeddings [pixel-dct]
img_dct_tile = self.nerf_image_embedder(nerf_pixels_tile)
Expand Down