sdxl-vae-fp16-fix / README.md
Seraph19's picture
Update README.md
fb510bc verified
|
raw
history blame
3.17 kB
metadata
license: mit
tags:
  - stable-diffusion
  - stable-diffusion-diffusers
inference: false
library_name: diffusers

SDXL-VAE-FP16-Fix

SDXL-VAE-FP16-Fix is the SDXL VAE*, but modified to run in fp16 precision without generating NaNs.

VAE Decoding in float32 / bfloat16 precision Decoding in float16 precision
SDXL-VAE ⚠️
SDXL-VAE-FP16-Fix

🧨 Diffusers Usage

Just load this checkpoint via AutoencoderKL:

import torch
from diffusers import DiffusionPipeline, AutoencoderKL

vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", vae=vae, torch_dtype=torch.float16, variant="fp16", use_safetensors=True)
pipe.to("cuda")

refiner = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-1.0", vae=vae, torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
refiner.to("cuda")

n_steps = 40
high_noise_frac = 0.7

prompt = "A majestic lion jumping from a big stone at night"

image = pipe(prompt=prompt, num_inference_steps=n_steps, denoising_end=high_noise_frac, output_type="latent").images
image = refiner(prompt=prompt, num_inference_steps=n_steps, denoising_start=high_noise_frac, image=image).images[0]
image

Automatic1111 Usage

  1. Download the fixed sdxl.vae.safetensors file
  2. Move this sdxl.vae.safetensors file into the webui folder under stable-diffusion-webui/models/VAE
  3. In your webui settings, select the fixed VAE you just added
  4. If you were using the --no-half-vae command line arg for SDXL (in webui-user.bat or wherever), you can now remove it

(Disclaimer - I haven't tested this, just aggregating various instructions I've seen elsewhere :P PRs to improve these instructions are welcomed!)

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.


* sdxl-vae-fp16-fix is specifically based on SDXL-VAE (0.9), but it works with SDXL 1.0 too