search

Magicboomliu / Accelerator-Simple-Template

This is a simple template using HuggingFace Accelerator for DDP-training/Saving/Loading/Pushing.
MIT License
48 stars 5 forks source link

image normaliztion #4

Closed wangchust closed 8 months ago

wangchust commented 8 months ago

left_image_data_resized in the training code is normalized to [0, 1] and fed to the vae encoder. However, according to Marigold, the vae encoder accepts data range [-1, -1]. Is this on purpose?

Magicboomliu commented 8 months ago

Since they are only the prompt, you can set to -1 to 1, but 0~1 is also Ok, it does not change the result a lot. Actually the inference code in the official repo, the prompt part(rgb latent)'s input is also 0-1

wangchust commented 8 months ago

Thanks for your response!