Ruoxi

Currently we only output those six views. Re-training is of course a valid solution.

Thank you for your detailed thoughts. I think it is possible and a good direction to work on.

Huggingface ran out of GPUs 😄. We have scheduled restart but can hardly do anything beyond that.

i have the same problem here; do you have any solutions? @davidvoisonics sorry for bumping but thanks in advance

I set `--invalidate-limit` to 32 or 64 but it doesn't help...

Transform-related stuff by design shall go into something like transform3d and that's why I did not include transform hierarchies, compositions or the tricky quaternion/rotation matrix/euler angles in the first place....

I tested that the scripts were working for step-based training. For epoch-based training I do think `num_update_steps_per_epoch` should be divided by the number of processes -- this value appears incorrect.

Thank you for your report! Could you specify a little bit more -- are you using jax for inference or you are using torch but importing diffusers causes a problem?

Problem still here after two years...disappointed

Do you have any idea how the orphans appear? It seems that from several days ago our seaweedfs starts to grow on disk rapidly (50GB/min) with nobody writing anything. It...