Better Colors VAE XL
Hello everyone! If you're reading this, you're probably not alone who noticed the original sdxl vae leaving white 'fog' on decoded latents. Well, I tried to change this and I think my experiment was sucessful.
Couldn't find good enough vae to use instead of the original one, so I decided to make my own. Now as a small Christmas gift for this year I present to you my 3 vae merge collection - Better Colors Vae XL in Vibrant, Perfect and Monumental releases. These are my attempts at making outputs from sdxl based models more colorful than ever. Yes, this is a merge - no training at all. I made it entirely in ComfyUI, showcase images as well. Workflow included.
Vibrant
This vae is made from original sdxl vae and another one making very dark images. Good enough still worth posting.
Perfect
This one more proper merge out of 7 other vaes. Less darker than Vibrant and still plenty colorful.
Monumental
The last vae is enhancement of Perfect version with another vae outputting decent sharp details when zooming very close at the output. My most reccomended version to use.
Note: This vae is not limited just to furry stuff. It should work with any sdxl based checkpoint, even Pony, not only Illustrious or NoobAI. But remember! Before baking to your model, test if it works like you expect. Don't use if it causes serious issues. If you know how to fix this, try making your own vae and let me know. Thanks!
And my last words - might produce dark outputs/change colors when inpainting. If it happens, try this one.
Credits
Thanks to all creators of vaes I used to mix my own:
Anzhc, Arctenox, atw44qb, easygoing0114, Karely_ai, konapieces, Liquidn2, Pitpe11, stabilityai
Description
FAQ
Comments (5)
Хороший VAE. Использовал версию Monumental. Получились яркие и насыщенные изображения с хорошей четкостью. Спасибо автору!
"who noticed the original sdxl vae leaving white 'fog' on decoded latents" No, more like it doesn't give enough saturation, it's extremely visible. I've been using custom VAEs
Didn't use VAEs from other creators like Felldude, ollin and etc :c
"Don't use if it causes serious issues. If you know how to fix this, try making your own vae and let me know" How and why would it cause trouble? It shouldn't. I made a script to merge VAEs, but others don't, so they don't know how to merge
I wish it could be easy to train VAEs or at least have a lot of merges, it is not limited to merging by blending or adding. Most don't care about autoencoders and go with the default ones (SDXL 0.9 VAE, SDXL VAE fp16 Fix) or choose a random one without knowing nor comparing to or trying others





