Ioannis Pikoulis

Figures 26, 27 in the appendix state: `Sampled with classifier-free guidance...` which is on par with the current code, i.e., the model is trained with class-context embeddings being fed through...

I'm currently conducting experiments with classifier-free guidance during training for the AffectNet dataset. I'm also considering whether the `null` embedding needs to be trainable or not. I will come back...

having the same problem. It has something to do with transformers package and pl version

if you remove the transformers package, then it runs correctly. Also kornia (0.5.0) needs to be installed and torchmetrics 0.10 or 0.11

> I get the same issue. I updated pytorch-lightning to a more recent version like 1.5.0, which fixed this error: `cannot import name 'get_num_classes' from 'torchmetrics.utilities.data'` but then it created...

> I have the same problem Downgrade torchmetrics to 0.5

> Hitting the same issue with `imagenet64_uncond_100M_1500K.pt` and 4k steps. Running with fewer diffusion steps (200, 500, 800, 1000) seems to work. > > I'll get back to the thread...

After using --diffusion_steps=1000 instead of 4000 for sampling from cifar10_uncond_50M_500K.pt & imagenet64_uncond_100M_1500K.pt, I started seeing clearer pictures instead of pure noise. I also tried using --diffusion_steps=4000 with the --timestep_respacing argument...

I also used image_nll.py for cifar10_uncond_50M_500K.pt. Using --diffusion_steps=4000 I got 5.25 bpd while using --diffusion_steps=1000 I got a better result of 3.28 bpd which is a lot closer to the...