Experiments on Diffusion

[Yeez-lee]

Hi,

It's a nice work. But I have questions about experiments on Diffusion.

1. In Table 8, do you compare your results with full data training (shown as original by 7.83)? But in E.2 VISUALIZATION OF DIFFUSION RESULT, you claim that the image quality is as good as trained with uniform random sampling, while the training cost is saved by 27%. How can I understand it?
2. When can you release your codes about this part?

[henryqin1997]

1. FID(Frechet Inception Distance) is lower the better, ours is 7.7. Our result is slightly better but this is not a big difference since our method is not targeting performance improvement.
2. Plan to release this part in Feb, keep an eye on the update.

[Yeez-lee]

Thanks for your help. But I want to clearify that in Table 8, you compare your results with full data training not uniform random sampling. That's right?

[henryqin1997]

It is full data training with uniform random sampling.