The Pulse
StableVQ Proposes Training Changes for More Stable Visual Tokenizers
The arXiv record contains no attributable quotation; I checked the title, author list, submission history and abstract.
AI.info Team ·
StableVQ is a paper about training stability in vector-quantized visual tokenizers, the systems that convert images into discrete visual representations for generative models. The paper was submitted to arXiv on September 22, 2026, under the title StableVQ: Practical Guidelines for Stable Vector-Quantized Tokenizer Training.
Its abstract says vector quantization is a foundation for discrete visual tokenizers used by modern autoregressive and masked image-generation models. It also says recent shared-projection codebook methods have improved codebook utilization, while training stability remains an unresolved challenge. The paper does not provide the numerical results, model comparisons or single-layer configuration described in the original article.
Three changes to the training process
The authors describe StableVQ as a method that revisits the learning objectives of the encoder-decoder and the codebook. Their stated concern is that the two parts of a tokenizer can become too dependent on one another during training. When each subsystem cannot reliably perform its own role, the overall system may work only when the two happen to cooperate. The abstract describes that condition as fragile, particularly when training is under stress.
StableVQ proposes three interventions. Dynamic STE changes the encoder’s learning objective by adjusting the behavior of the straight-through estimator. According to the abstract, this is intended to help the encoder optimize the reconstruction space under discrete regularization even when codebook utilization is low.
Region VQ Loss changes the codebook’s learning objective. The authors say it is designed to let the codebook track the encoder output distribution without depending on oscillations in the encoder to activate entries. The abstract presents this as a way to give the codebook a more independent training role.
The third component, Decoupled Schedule, assigns separate learning-rate schedules to the encoder-decoder and the codebook. The paper’s stated rationale is that the two modules have different responsibilities and therefore require different optimization dynamics.
Built on shared-projection codebooks
The abstract identifies StableVQ as a method built on shared-projection codebooks. It describes the approach as lightweight and says it introduces no learnable parameters. The source does not state that the method uses a single linear layer, and the arXiv abstract does not report specific reconstruction, perceptual-quality, generation, recall or ablation figures.
The paper says its experiments use ImageNet and demonstrate consistent improvements in training stability, codebook utilization and reconstruction quality across different codebook sizes and initialization settings. Those statements describe the direction of the reported findings, but the fetched arXiv record does not provide the underlying measurements or detailed experimental configurations.
What the source establishes
The available record establishes the paper’s title, submission date, authorship and central proposal. The listed authors are Bao Tang, Jiahao Guo, Haoxiang Cao, Wenyu Liu, Changqian Yu, Kun Gai and Xinggang Wang. The abstract frames StableVQ as a parameter-free set of training interventions for shared-projection visual-tokenizer codebooks.
It does not establish a 100% utilization result, a particular rFID or LPIPS score, comparisons with SimVQ or FVQ, a frozen-codebook pilot, model-size results, epoch counts, institutional affiliations or downstream generation metrics. Those details are therefore not included here.