StableVQ Reports 100% Codebook Use, rFID 1.05 on ImageNet

StableVQ, a tokenizer-training method from a team at Huazhong University of Science and Technology, KlingAI Research and South China Normal University, reports full codebook utilization with rFID 1.05 on ImageNet at a 262,144-code book, according to the paper on Hugging Face.

The authors trace Vector Quantization training failure to the "entanglement of the Encoder–Decoder and Codebook training." Their remedy is three parameter-free changes: a Dynamic Straight-Through Estimator that reweights encoder gradients by quantization error, a Region VQ Loss that propagates targets from active codes to inactive ones, and a decoupled learning-rate schedule running the codebook at a constant high rate while the encoder-decoder warms up and anneals.

On the paper's Usage Recovery AUC robustness measurement, StableVQ posts 60.59 ± 1.52 against 8.08 for FVQ and 2.17 for SimVQ. Against SimVQ at a 16,384-code book, reconstruction improves from 2.89 to 1.22 rFID with codebook usage still at 100%.

The abstract calls the method "lightweight" and says it "introduces no learnable parameters," with the reported gains holding "across diverse codebook sizes and initialization settings." It lands in a busy stretch for tokenizer and training-efficiency work in our open source AI feed.