PRQ-KMeans: Projection Residual Quantization for Semantic ID Tokenization
This work proposes PRQ-KMeans, which removes the global-mean component, refines centroids with Top-k similarity-weighted updates, and replaces full-codeword subtraction with a projection residual that removes each representation's selected-centroid component.
Yunxiao Luo, Siyuan Wang, Ben Chen et al.
· 0 citations