How to reproduce the results in the paper

[fierytree]

Hello, first of all, thank you very much for your fantastic work. Here I have some questions:

1. It seems that the file for calculating FVD and IS results is not provided in the code.
2. "diffusion_ddp_accum_constant_cu.py" is not provided. Is it the same as "diffusion_ddp_accum_constant_cu_frame_CLEVRER_robust.py"?
3. Is the preprocess of CLEVRER ("preprocess_CLEVRER_dataset.py") and other datasets the same? It would be very helpful if you can provide the train code and preprocess code of the other datasets (UCF, TaiChi, Sky).

[eecoder-dyf]

I have the same question, too. I run video_generation_CLEVRER.py with the pretrained model checkpoint_accum_clevrer_robust_400000.pth.tar and put one frame from CLEVRER dataset into folder frame_results , but the generated videos are still noise. Does anyone know how to use the pretrained model to generate results?