Skip to content

Discrepancy in Channel Dimensions: patch_embedding_extra (16 ch) vs. Wan Base Model (36 ch) #3

Description

@rain152

Hi authors, thank you for your impressive work on IC-Effect!

I noticed a potential dimension mismatch in models/transformer_wan.py.
The base model (Wan) typically operates on a 36-channel latent space (due to its specific VAE architecture). However, the newly added branch for conditioning, specifically the patch_embedding_extra layer, seems to be initialized or hardcoded with 16 channels.

Then I got: RuntimeError: The size of tensor a (36) must match the size of tensor b (16) at non-singleton dimension 1

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions