Reproduce “StableVQ: Practical Guidelines for Stable Vector-Quantized Tokenizer Training” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Bao Tang, Jiahao Guo, Haoxiang Cao, Wenyu Liu, Changqian Yu, Kun Gai, Xinggang Wang

StableVQ improves training stability in image tokenizers by assigning separate roles to the encoder–decoder and codebook: Dynamic STE moderates unreliable reconstruction gradients, Region VQ guides inactive codes, and a Decoupled Schedule uses different learning-rate schedules. On controlled initialization mismatches, it achieves UR-AUC 60.59 ± 1.52 versus FVQ’s 8.08 ± 0.24, for tested settings only.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.