File size: 4,960 Bytes
b7e7887 2869c9a b7e7887 2869c9a b7e7887 2869c9a b7e7887 2869c9a b7e7887 2869c9a b7e7887 2869c9a b7e7887 2869c9a b7e7887 |
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 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 |
# Improved Autoencoders
## Decoder Finetuning
We publish two kl-f8 autoencoder versions, finetuned from the original [kl-f8 autoencoder](https://github.com/CompVis/latent-diffusion#pretrained-autoencoding-models).
The first, _ft-EMA_, was resumed from the original checkpoint, trained for 313198 steps and uses EMA weights.
The second, _ft-MSE_, was resumed from _ft-EMA_ and uses EMA weights and was trained for another 280k steps using a re-weighted loss, with more emphasis
on MSE reconstruction (producing somewhat ``smoother'' outputs).
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/stable-diffusion-decoder-finetune/blob/main/ckpt_00560000-kat-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/stable-diffusion-decoder-finetune/blob/main/ckpt_00840000-kat-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/stable-diffusion-decoder-finetune/blob/main/ckpt_00560000-kat-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/stable-diffusion-decoder-finetune/blob/main/ckpt_00840000-kat-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._
<p align="center">
<br>
<b>
256x256: ft-EMA (left), ft-MSE (middle), original (right)</b>
</p>
<p align="center">
<img src=eval/ae-decoder-tuning-reconstructions/merged/00025_merged.png />
</p>
<p align="center">
<img src=eval/ae-decoder-tuning-reconstructions/merged/00011_merged.png />
</p>
<p align="center">
<img src=eval/ae-decoder-tuning-reconstructions/merged/00037_merged.png />
</p>
<p align="center">
<img src=eval/ae-decoder-tuning-reconstructions/merged/00043_merged.png />
</p>
<p align="center">
<img src=eval/ae-decoder-tuning-reconstructions/merged/00053_merged.png />
</p>
<p align="center">
<img src=eval/ae-decoder-tuning-reconstructions/merged/00029_merged.png />
</p>
|