Opensora 1.2 inference fail

[qian18long]

The inference video is only noise from the given checkpoint: https://huggingface.co/LanguageBind/Open-Sora-Plan-v1.2.0

[TianxiangMa]

Me too, why?

0.1.mp4

[ytm-01]

i meet the same question.

[TianxiangMa]

I solved this problem by adjusting the "num_frames" to 93. The trained 4-second model can only be used to inference 4-second videos.

[LinB203]

Dynamic length of time is still in the training is not released for the time being.

[LinB203]

Btw, our project name is Open-Sora Plan, not Open-Sora, which is an another repo.

[foreverpiano]

@LinB203 想问下现在inference时候，DiT 的架构3D是怎么理解的？以前是 spatial temporal 交替，然后把多余的 reshape 到 bs 的维度，现在我看代码是只有一个 transformer 了，现在还是 transformer2D 吗？求解释下架构