README.md

---
datasets:
- Disty0/danbooru_curated-jxl_lossless_4mp
base_model:
- stabilityai/stable-diffusion-3.5-medium
- stabilityai/stable-diffusion-3-medium-diffusers
pipeline_tag: text-to-image
library_name: diffusers
---

This VAE was finetuned on PNG only anime illustrations for 512 steps with MAE Loss.  
Used fp32 weights + fp16 mixed precision with learning rate 4e-6 and effective batch size of 16.  

This training was to test my VAE decoder training code and 512 step model turned out to be better than i expected and fixes the color shifting issues of the original SD3 VAE pretty well.  
I stopped messing with SD3 after a while but i decided to release this VAE finetune instead of deleting it.  



Original Image:

![orig.png](https://cdn-uploads.huggingface.co/production/uploads/6456af6195082f722d178522/MEGBwJ5wyjS4sGa1otHDW.png)


Original SD3 VAE:

![vae_orig.png](https://cdn-uploads.huggingface.co/production/uploads/6456af6195082f722d178522/_QWnQuYJ7BaJPyaVpb6gq.png)

Anime VAE Finetune:


![vae_ft.png](https://cdn-uploads.huggingface.co/production/uploads/6456af6195082f722d178522/5q7goj2J2TjziPP2iCVT_.png)