Fixed VAE is very low quality

replicate / cog-sdxl

Stable Diffusion XL training and inference as a cog model

https://replicate.com/stability-ai/sdxl

Apache License 2.0

212 stars 106 forks source link

Fixed VAE is very low quality #49

Open levi opened 9 months ago

levi commented 9 months ago

Using the built-in upcast fixed vae, fine details like hair come out fuzzy and pixelated. This is particularly present if cog-sdxl is used to generate a full fine tune and then loras are created from the full fine tune from images generated from the model.

Changing to use madebyollin/sdxl-vae-fp16-fix for fp16/bf16 training fixes the issue.

Generation from full fine tune:

Generation from LoRA trained on full fine tune as base model:

zeke commented 8 months ago

Thanks for the heads up, @levi. Would you be able to open a PR to address this?