Hyper-Spherical Quantization for Discrete Visual Representations

A recent preprint on arXiv (2607.22148) presents Hyper-Spherical Quantization (HSQ) as a solution to the issue of codebook collapse in the discretization of high-dimensional visual data. The authors pinpoint the source of the problem as a metric mismatch between typical Euclidean objectives and the anisotropic nature of representation space, which results in inconsistent angular distributions and high-variance magnitude scales. HSQ utilizes angular routing to separate semantic content from feature magnitude, thus avoiding code assignments that are overly influenced by scale. Consequently, the discrete Representation Autoencoder (dRAE) is able to achieve high-quality reconstruction while maintaining semantic coherence and accommodating scalable codebook budgets.