Aaah, vibrant VAE i guess

This is a by-product of attempts to finetune SDXL vae.
It does not change reconstruction significantly, but has higher vibrancy, depending on generation. Especially vibrant when using hires.

This has no practical use in your general generations, unless you are trying to replicate toxicly vibrant styles, or just love when your eyes bleed out when they see bright color.

Essentially, it seem to extend dynamic range.

It can slightly alter generations in certain cases, but overall it is roughly the same.

P.S. Those colors are not more correct, or better, it's up to you. This just adds vibrancy and sometimes contrast.

How did i make it

I was about to go crazy due to models not operating properly after training. Then i took KL divergence loss out and everything worked.

This model is a by-product of testing and it utilized custom channel normalization, which i think is not supposed to be used in normal VAE training in general, but here are results of that.

Training took like 8 minutes i think. Tuned directly from SDXL vae on small anime dataset.