[bugfix] Wan I2V: CLIP image conditioning silently dropped when passed as a tensor during training

[shaoxiongduan]

Purpose

WanTransformer3DModel.forward (and the causal/dreamx/lingbotworld variants) unwrapped encoder_hidden_states_image only when it was a list, and set it to None otherwise — discarding a bare tensor even though the signature accepts torch.Tensor | list[torch.Tensor] | None.

The inference pipelines store image_embeds as a list (image_encoding.py), so they were unaffected. Every I2V training pipeline passes the CLIP features directly as a tensor (wan_i2v_training_pipeline.py:129, wan_i2v_distillation_pipeline.py:170, self_forcing_distillation_pipeline.py), so those trained with no image cross-attention at all, silently and without error.

Models with image_dim=None (Wan2.2 I2V/TI2V, LingBotWorld) build no image embedder and are unaffected either way.

Changes

Unwrap a list, leave a tensor alone. Matches the handling of encoder_hidden_states two lines above, and the form already used in matrixgame2/causal_model.py and matrixgame2/model.py.

Touches wanvideo.py, causal_wanvideo.py (2 sites), dreamx_world.py, lingbotworld/model.py.

Test Plan

Wan2.1-I2V-14B-720P, one image, 480x832, 121 frames, 30 steps, guidance 3.0, seed 1000.

The inference stack passes a list, so it never hits the bug. To exercise the training convention end-to-end, denoising.py was temporarily patched to hand the DiT a bare tensor — this patch is not part of the PR:

# DenoisingStage.forward, right after:
#   image_embeds = [image_embed.to(target_dtype) for image_embed in image_embeds]
image_embeds = image_embeds[0]   # training convention: bare tensor

Then generate twice on the same node/seed: once with wanvideo.py from main, once with this branch.

Test Results

std = per-video pixel standard deviation (contrast/detail proxy).

	mean	std	last-frame std
before — tensor silently nulled	120.1	60.5	57.5
after — tensor kept	80.7	63.7	66.3
unmodified inference (list, reference)	80.7	63.8	66.8

After the fix, passing a tensor reproduces the untouched list path exactly. Before it, the output is brighter and washed out, degrading over the clip.

Same fix measured through a training-side validation callback (Wan2.1-I2V-14B, stock weights), against ground truth:

clip	before	after	GT
720p, clip A	19.3	70.3	64.7
720p, clip B	26.8	57.2	58.4
480p, clip A	32.0	70.3	64.6

Checklist

• pre-commit run — all four files live under fastvideo/models/, which the config's global exclude skips, so no hooks apply to this diff
• I added or updated tests for my changes — none added. A regression test asserting forward(..., encoder_hidden_states_image=tensor) != forward(..., encoder_hidden_states_image=None) would prevent this from recurring; happy to add if wanted.
• I updated documentation if needed — n/a
• I considered GPU memory impact — none

For model/pipeline changes, also check:

• I verified SSIM regression tests pass — not run locally; inference output is unchanged (list path untouched), so SSIM should be unaffected
• I updated the support matrix — n/a

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[mergify]

Merge Protections

🔴 1 of 1 protections blocking · waiting on 👀 reviews and 🤖 CI

	Protection	Waiting on
🔴	PR merge requirements	👀 reviews and 🤖 CI

🔴 PR merge requirements

Waiting for

• #approved-reviews-by>=1
• check-success=fastcheck-passed
• check-success=full-suite-passed

This rule is failing.

• #approved-reviews-by>=1
• check-success=fastcheck-passed
• check-success=full-suite-passed
• check-success~=pre-commit
• title~=(?i)^\[(feat|feature|bugfix|fix|refactor|perf|ci|doc|docs|misc|chore|kernel|new.?model|skill|skills|infra)\]

[shaoxiongduan]

/merge