Edit model card

These VAEs are modified versions of madebyollin's sdxl-vae-fp16-fix.

These VAEs should not produce a NaN in VAE error even when used on half precision.

They have been modified using the ideas from VAE-BlessUp script to produce higher contrast and lower brightness images than the original version.

The recommended version is sdxl-vae-fp16fix-blessed.safetensors

For most SDXL models, you probably should probably just use the non-blessed sdxl-vae-fp16-fix.

I made these mostly for fun, but found that slightly increasing contrast and decreasing brightness actually improved the outputs on the model I was testing.

You may find one of them to be beneficial for PonyDiffusionV6-XL and other models based on it.

Best - sdxl-vae-fp16fix-blessed.safetensors = 1.1 contrast multiplier/0.7 brightness multiplier

Good - sdxl_vae-fp16fix-c-1.1-b-0.5.safetensors = 1.1 contrast multiplier/0.5 brightness multiplier

High Contrast - sdxl_vae-fp16fix-c-1.2-b-0.7.safetensors = 1.2 contrast multiplier/0.7 brightness multiplier

Very High Contrast - sdxl_vae-fp16fix-c-1.2-b-0.5.safetensors = 1.2 contrast multiplier/0.5 brightness multiplier

Untested:

sdxl_vae-fp16fix-c-0.9.safetensors = 0.9 contrast multiplier

sdxl_vae-fp16fix-c-0.9-b-0.9.safetensors = 0.9 contrast multiplier/0.9 brightness multiplier

sdxl_vae-fp16fix-c-0.8.safetensors = 0.8 contrast multiplier

sdxl_vae-fp16fix-c-0.8-b-0.9.safetensors = 0.8 contrast multiplier/0.9 brightness multiplier

sdxl_vae-fp16fix-c-0.8-b-0.8.safetensors = 0.8 contrast multiplier/0.8 brightness multiplier

Example images (made using AutismMix_confetti):

Thank you Neggles for the script used to make them!

Downloads last month
65,626
Inference API
Unable to determine this model’s pipeline type. Check the docs .

Spaces using nubby/blessed-sdxl-vae-fp16-fix 16