ArcVQ-VAE: A Spherical Vector Quantization Framework with ArcCosine Additive Margin

Researchers introduced ArcVQ-VAE, a VQ-VAE variant that adds spherical codebook constraints and an angular-margin loss.

The method combines ball-bounded norm regularization with an ArcCosine Additive Margin Loss to shape the codebook space.

The goal is to overcome codebook capacity limits, improve latent coverage, and make discrete tokens more efficient.

The paper reports competitive results on image reconstruction and generation, with better representation diversity and sample quality.