StableVQ separates encoder and codebook training to stabilise image tokenizers

single source· 1 articles · confidence: medium · first seen 2026-09-21 20:00 UTC

What this means for you

Nothing to act on unless you train vector-quantised tokenizers yourself. The abstract names no code release and reports no scores, so there is nothing to reproduce. If you do train them, this is cheap to try later: the changes add no learnable parameters.

StableVQ is a training recipe for vector-quantised image tokenizers — the component that converts a picture into a short sequence of dictionary entries so an image generator can predict it the way a language model predicts words. The paper argues the instability practitioners hit is structural: the encoder-decoder and the codebook (that shared dictionary) are trained together, and neither can do its job alone, so the system holds only while the two happen to cooperate. Three changes are proposed: a revised straight-through estimator, a region-based codebook loss, and independent learning-rate schedules for each half. It adds no learnable parameters. ImageNet experiments report better stability, dictionary utilisation and reconstruction; no scores are given in the abstract.

Key facts

  • ·StableVQ is a training method for vector-quantised (VQ) image tokenizers, posted as arXiv preprint 2609.26774 on 21 September 2026. source
  • ·The method has three components: Dynamic STE, Region VQ Loss and Decoupled Schedule. source
  • ·The paper attributes training instability to entanglement between encoder-decoder training and codebook training. source
  • ·StableVQ builds on shared-projection codebooks, adds no learnable parameters, and is described as lightweight. source
  • ·Experiments on ImageNet report improvements in training stability, codebook utilisation and reconstruction quality across codebook sizes and initialisation settings; the abstract states no numeric scores. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire