File size: 1,023 Bytes
c66f1b8 38c3c49 c66f1b8 38c3c49 5a7a4f5 38c3c49 5a7a4f5 38c3c49 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 |
---
license: openrail++
tags:
- stable-diffusion
inference: false
---
this is merge model for:
1. 100% stable-diffusion-xl-base-1.0
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
2. sdxl-vae-fp16-fix
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
you can use this directly or finetune.
same license on stable-diffusion-xl-base-1.0
same vae license on sdxl-vae-fp16-fix
# SDXL-VAE-FP16-Fix
SDXL-VAE-FP16-Fix is the [SDXL VAE](https://huggingface.co/stabilityai/sdxl-vae)*, but modified to run in fp16 precision without generating NaNs.
## Details
SDXL-VAE generates NaNs in fp16 because the internal activation values are too big:
SDXL-VAE-FP16-Fix was created by finetuning the SDXL-VAE to:
1. keep the final output the same, but
2. make the internal activation values smaller, by
3. scaling down weights and biases within the network
There are slight discrepancies between the output of SDXL-VAE-FP16-Fix and SDXL-VAE, but the decoded images should be close enough for most purposes.
--- |