Binary VAE with Continuous Channels

This demo shows how the reconstruction quality improves as we use more continuous latent channels. The model uses 256 discrete (ternary) channels + 0-32 continuous channels. Input images are resized so their shortest edge is 512, and then cropped so dimensions are perfectly divisible by 32.

Channel counts tested: 0, 8, 16, 32

Input Image

Original (preprocessed)

Reconstructions (2x2 Grid)

Top row: 0ch, 8ch | Bottom row: 16ch, 32ch

Reconstructions Grid

Metrics

Notes:

PSNR: Peak Signal-to-Noise Ratio (higher is better, >30 dB is good)
SSIM: Structural Similarity Index (0-1, higher is better)
The model has 32 continuous channels max.
Discrete latent (256 ternary channels) is always active.