Failed to reproduce FVD reported in the paper #65
Description
@wyhsirius @maxin-cn @xinyuanc91 @pooya-mohammadi I'm trying to reproduce zero-shot UCF101 FVD score of LaVie (526.3) reported in Table 2. However, I'm only getting much higher FVD score (689.6), and I hope you could help me on this.
Below are some details of my trial:
- For video generation using LaVie, I followed all details presented in section 5.4 (class name as text prompt, 100 videos per class, etc..)
- For sampling config, although it is not presented in the paper, I used base/configs/sample.yaml
- For FVD computation, I utilized the code from this repo.
Where do you think the score difference comes from? Some missing information that I think might cause the difference are:
- Sampling config: (num_sampling_steps, cfg_scale, ...) Did you use the ones in base/configs/sample.yaml ?
- FVD computation code: Which codebase did you use for FVD computation?
- real video stats: I randomly sampled 10k videos from whole UCF101 dataset. How exactly did you compute real video stats?
If you can find anything that I'm missing or doing incorrectly here, please let me know.
Thank you.