license: openrail
library_name: diffusers
EQ-Adapted Noobai11 Unet
Using B3/B4/B5 SDXL EQ VAE from here: https://huggingface.co/Anzhc/MS-LC-EQ-D-VR_VAE
You need Reflect pad in conv layers of VAE to properly use it(otherwise there will be some artefacts on edges), refer to VAE repo for instructions. If you have trained without this - don't worry, just add modification and do a small tune to fix up artefacts on edges.
This is an EQ adapted unet, that learned to denoise clean latent representations, as opposed to noised default SDXL VAE representation.
This halves loss in training. Yep.
In my tests it also improves convergence, and makes model more trainable, or so it seems at the very least.
Additionally, this alleviates some of the issues of Noob11 base, like overly bright images.
Some more graphs for more technical people:
