Residual VQ

Residual vector quantization (RVQ) approximates a vector or signal by adding codebook entries in stages, with each stage encoding the error left by the previous ones. Researchers use it in neural audio codecs and generative tokenizers for images, motion, and physiological signals, as well as for recommendation IDs, retrieval, and compressed model caches. Common approaches combine RVQ with autoencoders, coarse-to-fine autoregressive prediction, progressive coding, or task-aware objectives, and are evaluated on reconstruction quality, bitrate, latency, and downstream tasks. A central tension is balancing compact codes and efficient decoding against fidelity and useful semantic structure; stage-aware geometry, adaptive codebooks, and unequal error costs remain active design challenges.