sd-vae-ft-ema / README.md
patrickvonplaten's picture
Remove text to image tag
f04b2c4
metadata
license: mit
tags:
  - stable-diffusion
  - stable-diffusion-diffusers
inference: false

Improved Autoencoders

Utilizing

These weights are intended to be used with the 🧨 diffusers library. If you are looking for the model to use with the original CompVis Stable Diffusion codebase, come here.

How to use with 🧨 diffusers

You can integrate this fine-tuned VAE decoder to your existing diffusers workflows, by including a vae argument to the StableDiffusionPipeline

from diffusers.models import AutoencoderKL
from diffusers import StableDiffusionPipeline

model = "CompVis/stable-diffusion-v1-4"
vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-ema")
pipe = StableDiffusionPipeline.from_pretrained(model, vae=vae)

Decoder Finetuning

We publish two kl-f8 autoencoder versions, finetuned from the original kl-f8 autoencoder on a 1:1 ratio of LAION-Aesthetics and LAION-Humans, an unreleased subset containing only SFW images of humans. The intent was to fine-tune on the Stable Diffusion training set (the autoencoder was originally trained on OpenImages) but also enrich the dataset with images of humans to improve the reconstruction of faces. The first, ft-EMA, was resumed from the original checkpoint, trained for 313198 steps and uses EMA weights. It uses the same loss configuration as the original checkpoint (L1 + LPIPS). The second, ft-MSE, was resumed from ft-EMA and uses EMA weights and was trained for another 280k steps using a different loss, with more emphasis on MSE reconstruction (MSE + 0.1 * LPIPS). It produces somewhat ``smoother'' outputs. The batch size for both versions was 192 (16 A100s, batch size 12 per GPU). To keep compatibility with existing models, only the decoder part was finetuned; the checkpoints can be used as a drop-in replacement for the existing autoencoder.

Original kl-f8 VAE vs f8-ft-EMA vs f8-ft-MSE

Evaluation

COCO 2017 (256x256, val, 5000 images)

Model train steps rFID PSNR SSIM PSIM Link Comments
original 246803 4.99 23.4 +/- 3.8 0.69 +/- 0.14 1.01 +/- 0.28 https://ommer-lab.com/files/latent-diffusion/kl-f8.zip as used in SD
ft-EMA 560001 4.42 23.8 +/- 3.9 0.69 +/- 0.13 0.96 +/- 0.27 https://huggingface.co/stabilityai/sd-vae-ft-ema-original/resolve/main/vae-ft-ema-560000-ema-pruned.ckpt slightly better overall, with EMA
ft-MSE 840001 4.70 24.5 +/- 3.7 0.71 +/- 0.13 0.92 +/- 0.27 https://huggingface.co/stabilityai/sd-vae-ft-mse-original/resolve/main/vae-ft-mse-840000-ema-pruned.ckpt resumed with EMA from ft-EMA, emphasis on MSE (rec. loss = MSE + 0.1 * LPIPS), smoother outputs

LAION-Aesthetics 5+ (256x256, subset, 10000 images)

Model train steps rFID PSNR SSIM PSIM Link Comments
original 246803 2.61 26.0 +/- 4.4 0.81 +/- 0.12 0.75 +/- 0.36 https://ommer-lab.com/files/latent-diffusion/kl-f8.zip as used in SD
ft-EMA 560001 1.77 26.7 +/- 4.8 0.82 +/- 0.12 0.67 +/- 0.34 https://huggingface.co/stabilityai/sd-vae-ft-ema-original/resolve/main/vae-ft-ema-560000-ema-pruned.ckpt slightly better overall, with EMA
ft-MSE 840001 1.88 27.3 +/- 4.7 0.83 +/- 0.11 0.65 +/- 0.34 https://huggingface.co/stabilityai/sd-vae-ft-mse-original/resolve/main/vae-ft-mse-840000-ema-pruned.ckpt resumed with EMA from ft-EMA, emphasis on MSE (rec. loss = MSE + 0.1 * LPIPS), smoother outputs

Visual

Visualization of reconstructions on 256x256 images from the COCO2017 validation dataset.


256x256: ft-EMA (left), ft-MSE (middle), original (right)