Generation FID with CFG in 80 epochs and 400 or 800 epochs

[Tom-zgt]

Great work!
Thanks a lot for sharing this impressive repository.
@JingfengYao
I have a few questions regarding the experimental results. I noticed that the paper doesn't explicitly report the following metrics for VTP + lightningDiT:

At 80 epochs: The FID score with Classifier-Free Guidance (w/ CFG).

At 400, 800, or more epochs: The FID scores both with (w/ CFG) and without (w/o CFG) Classifier-Free Guidance.

Could you kindly share these metrics if available? These details would be extremely helpful for me to better reproduce and follow your work.

Thanks again for your contribution!

[JingfengYao]

Hi,

Thanks for your interest. To be honest, we haven't been focusing on testing these specific metrics. We think ImageNet FID is reaching saturation, which is why we only reported results under limited computation or trend lines.

While we plan to evaluate VTP on larger-scale generation tasks in the future, we don't currently have plans to report improved ImageNet results in the short term.