license: mit
tags:

• stable-diffusion
• stable-diffusion-diffusers
• text-to-image
  inference: false

Improved Autoencoders

Utilizing

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

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)

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

Visual

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

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