← Return to Internet Discotheque

NEON VAE LAB // 32-32

IMAGE PLAYGROUND

A traced diffusion UNet compresses uploads into a 32x32 latent bottleneck, a residual MLP cleans the grid, and a VAE decoder pulls it back into RGB. DF2K super-resolution handles the HD pass, so it doesn't aim for strict photorealism-the model generalizes interesting textures, silhouettes, and color runs instead.

  • ✦ 32x32 latent stage feeding a traced diffusion UNet encoder/decoder
  • ✦ Residual MLP refiner keeps the latent grid steady before reconstruction
  • ✦ DF2K-trained upsampler handles the HD pass while preserving those non-photoreal patterns
  • ✦ Real-time processing, visualization, and transformation pipeline
  • ✦ Server-side inference runs full diffusion chain on CPU—processing may take 30-90s per job depending on load
TERMINAL DISCONNECTED
Click "Launch Pipeline" to connect
× Enlarged view