Venkat Krishnan

Results 10 comments of Venkat Krishnan

converted models using convert_original_stable_diffusion_to_diffusers.py are not identical

I'm having the same issue - any updates?

Outpus by trained model is very different from origin training materials

Hi, I was wondering if you had an update on if you were able to get better results?

Outpus by trained model is very different from origin training materials

Would using stable_txt2img.py instead of txt2img.py and SD embeddings give better results?

Seed not giving reproducibility

Yes, I believe so.

Seed not giving reproducibility

I also tried using seed_everything form the pytorch library but that still isn't producing reproducible models. Could there be variability between machines.

Seed not giving reproducibility

Thank you for the fast response, can you elaborate on this? I am training both the text encoder and unet together and specifying a single --seed argument to the train_dreambooth.py...

Seed not giving reproducibility

Looking at the colab code, the seed value from the box is passed into the --seed parameter of train_dreambooth.py as $Seed. The only difference I see between the way I'm...

Seed not giving reproducibility

Could this be related to https://pytorch.org/docs/stable/notes/randomness.html? I'm running these across different A10G's on AWS and have this issue. If I run repeatedly on single A10G it seems fine.

Same train seed produces different results

I think it is because Pytorch can be non deterministic: https://pytorch.org/docs/stable/notes/randomness.html

How much memory does this take for text_enc + unet with prior preservation?

For some reason when I run it on my custom computer with the same flags, it takes 24 GB. Do you know any settings or missing installations that could case...