dRAE Scales Visual Discretization to 130k Tokens

The newly proposed dRAE introduces a discrete representation autoencoder based on hyper-spherical quantization, scaling visual discretization to 131,000 tokens while effectively avoiding codebook collapse.

2026-07-28 ~ 2026-07-28 · 2 related posts

• dRAE scales visual tokenization to 131,072 codes without codebook collapse — burny_tech · 2026-07-28
• dRAE uses hyperspherical codes to scale visual discretization to 131,072 entries — antgroup · 2026-07-28