×ばつ Acceleration for Video Diffusion Models - Fix Wan JVP image projector checkpoint compatibility by zjn20030811 · Pull Request #140 · thu-ml/TurboDiffusion">
Skip to content

Navigation Menu

Sign in
Sign up

Fix Wan JVP image projector checkpoint compatibility - #140

Open
zjn20030811 wants to merge 1 commit into
thu-ml:main from
zjn20030811:fix/wan-jvp-mlp-layernorm
Open

Fix Wan JVP image projector checkpoint compatibility #140
zjn20030811 wants to merge 1 commit into
thu-ml:main from
zjn20030811:fix/wan-jvp-mlp-layernorm

Conversation

@zjn20030811

@zjn20030811 zjn20030811 commented Sep 3, 2026

Copy link
Copy Markdown

Summary

  • restore the Wan image-conditioning projector architecture used by the published I2V checkpoints
  • retain affine parameters and the canonical epsilon for both LayerNorm stages in the JVP implementation
  • add CPU regression coverage for checkpoint keys, forward equivalence, gradients, and first/last-frame token layout

Why

The JVP implementation constructed the projector with the model's non-affine WanLayerNorm. As a result, pretrained projector normalization weights and biases were dropped during loading, changing the image-conditioning transform. The projector now mirrors the canonical Wan I2V module, so its state dictionary and numerical path are compatible with those checkpoints.

Validation

python -m pytest -q tests/test_wan2pt1_jvp_mlp_proj.py
python -m compileall -q turbodiffusion tests

All three regression tests pass.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

AltStyle によって変換されたページ (->オリジナル) /