dRAE: Representation Autoencoder with Hyper-Spherical Codes

Researchers introduced dRAE, a discrete representation autoencoder for visual features.

It uses hyper-spherical quantization to separate semantic content from feature magnitude and avoid the codebook collapse seen in Euclidean quantization.

The method improves reconstruction, keeps 100% codebook usage, and scales to a vocabulary of 131,072.

It delivers strong results on understanding and generation tasks, pointing to scalable visual codes for multimodal AI systems.