The generated result is black，what's wrong with it?

[TZYSJTU]

hi，the demo result looks strange，it is all black
Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of .mp4
https://github.com/user-attachments/assets/cb953284-a617-43d4-980a-27f0eb682370

[TZYSJTU]

hi，the demo result looks strange，it is all black Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of .mp4 https://github.com/user-attachments/assets/cb953284-a617-43d4-980a-27f0eb682370

H20

[TZYSJTU]

RuntimeWarning: invalid value encountered in cast
outputs.append((x * 255).numpy().astype(np.uint8))
It seems that the results are NAN, then transferred to black

[TZYSJTU]

When I used --precision bf16， the generated results look right

A.lone.hiker.stands.atop.a.towering.cliff.silhouetted.against.the.vast.horizon.The.rugged.landscap.mp4

[foreverpiano]

Thanks for reporting this issue. The black output is caused by NaN values in computation getting cast to zeros. Using --precision bf16 resolves this numerical instability problem, as you found. Thank you for sharing this solution.

[zhisbug]

@TZYSJTU we'll support a lower precision soon, stay tuned.

[TimPietrusky]

@TZYSJTU I'm also trying to get this to work, which model did you use? https://huggingface.co/FastVideo/FastHunyuan or https://huggingface.co/FastVideo/FastHunyuan-diffusers?

I was using diffusers in combination with scripts/inference_diffusers_hunyuan.sh, but all results are black, even when i change the precision parameter.

What kind of setup do you have?

@foreverpiano or is the diffusers pipeline just not working right now?

[foreverpiano]

cc @rlsu9

[rlsu9]

@TimPietrusky Hi, could you please provide your scripts/inference_diffusers_hunyuan.sh? Thanks a lot.

[TimPietrusky]

@rlsu9 it's the one from this repository: https://github.com/hao-ai-lab/FastVideo/blob/main/scripts/inference/inference_diffusers_hunyuan.sh

[dabeschte]

@TimPietrusky the token refiner produces NaNs when using compiled. At least with torch 2.5.0 which is installed in env_setup.sh.

disabling compilation of the token refiner solves it.
add @torch.compiler.disable() in line fastvideo/models/hunyuan/modules/token_refiner.py:181 just before the forward function

Not sure if you used compilation as well. I only tried the inference_hunyuan_STA.sh script