BlinkDL/

clip-guided-binary-autoencoder

This model can encode 224x224 RGB image into 28x28x13bit (1274 bytes) latent. The compression rate is 28x28x13/(224x224x24)=1/118, or 0.203 bpp (same as VQGAN_f8_8192).

Demo: https://huggingface.co/spaces/Blealtan/clip-guided-binary-autoencoder

12M params for Encoder + Decoder. Trained on LAION-Aesthetics V2 5+ for 238M images.

Update: Now with 50M and 200M params checkpoints too :) Check the files.

Guided by https://huggingface.co/laion/CLIP-ViT-B-32-laion2B-s34B-b79K (it's great. better than OpenAI CLIP B/32) and https://github.com/dingkeyan93/DISTS. No GAN loss.