nodes/ARM64-EC-ComfyUI-LongCatPlugin
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix: 'Wrong Latent'

Browse Source
This commit is contained in:
ARM64_EC
2025-12-09 16:06:09 -05:00
parent d6cf1098b8
commit b3e274e4de
1 changed files with 33 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
nodes.py
+33 -9
View File
@@ -464,20 +464,44 @@ class LongCatSampler:
vae = neg_dict["vae"]
if images is not None:
from diffusers.image_processor import VaeImageProcessor
image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor * 2)
pil_imgs = tensor2pil(images[0])
processed_imgs = image_processor.resize(pil_imgs, height, width)
# Convert to tensor for ComfyUI VAE [B, H, W, C]
vae_input = pil2tensor(processed_imgs).to(device)
# Resize images manually to ensure correct dimensions
# pil2tensor returns [B, H, W, C]
pixel_images = pil2tensor(images[0]).permute(0, 3, 1, 2).to(device) # [B, C, H, W]
pixel_images = torch.nn.functional.interpolate(pixel_images, size=(height, width), mode="bicubic", align_corners=False)
vae_input = pixel_images.permute(0, 2, 3, 1) # [B, H, W, C]
# Encode using ComfyUI VAE wrapper
encoded = vae.encode(vae_input)
encoded = encoded.to(dtype=transformer.dtype)
# Check for channel mismatch (Flux VAE should be 16 channels, SD is 4)
if encoded.shape[1] == 4 and num_channels_latents == 16:
# If user provided a standard SD VAE (4ch) but model needs Flux VAE (16ch),
# we can try to adapt by pixel_unshuffle (4ch -> 16ch)
# [B, 4, H, W] -> [B, 16, H/2, W/2]
encoded = torch.nn.functional.pixel_unshuffle(encoded, 2)
# Now shape is [B, 16, H/2, W/2]
# But _pack_latents expects [B, 16, H, W] (full resolution)?
# No, _pack_latents expects the input to be viewable as [B, 16, H/2, 2, W/2, 2]
# which means it expects [B, 16, H, W].
# Wait, if we unshuffle, we reduce resolution.
# If we pass this reduced resolution to _pack_latents?
# _pack_latents will try to reduce it AGAIN.
# If the model is Flux, it expects 16 channel input to _pack_latents.
# This implies the VAE output is 16 channels at resolution H, W.
# If we have 4 channels at H, W.
# We cannot magically get 16 channels at H, W.
# Unless... the VAE output IS 16 channels.
# So the user MUST use a Flux VAE.
pass
if encoded.shape[1] != num_channels_latents:
raise ValueError(f"VAE output channels ({encoded.shape[1]}) does not match model expectation ({num_channels_latents}). Please ensure you are using a Flux VAE (16 channels).")
image_latents = _pack_latents(encoded, encoded.shape[0], num_channels_latents, latent_height, latent_width)
image_latents_ids = prepare_pos_ids(modality_id=2,