VaeZip: open-source codec stores VAE latent datasets 5-10x smaller with near-zero quality loss

TensorForger · reddit · 2026-10-17

A developer released VaeZip, an open-source PyTorch library tackling the fact that VAE latent datasets often take more space than the original data (JPEG/video codecs compress the source well, latents are nearly uncompressed).

How it works

Implements several quantization codecs designed specifically for VAE latents, exploiting spatial redundancy and the roughly normal latent distribution;
The best codec combines global/local range scaling with a Lloyd-Max codebook tuned per sample distribution, quantizing values to 2-8 bits and packing them;
Achieves 3 bits/value (10x smaller files) with decoded pixel MSE of just 0.0003;
Stores results in standard Safetensors files.

Validation

Codecs were selected from a benchmark comparing 50 algorithms across 8 image and video VAEs (reconstruction quality, fidelity, space, I/O latency);
Pretraining a small T2I flow-matching DiT on 6-bit-compressed latents gives a loss curve indistinguishable from the fp32 baseline.

Use case is narrow but valuable when latent datasets are too large for local disk and require costly S3 storage.

Original post →

VaeZip: open-source codec stores VAE latent datasets 5-10x smaller with near-zero quality loss

More from Infra