zhaoyue-zephyrus/InfinityCC_L24SQ

Spherical Leech Quantization for Visual Tokenization and Generation

     

This model implements Spherical Leech Quantization ($\Lambda_{24}$-SQ), a novel non-parametric quantization method for visual tokenization and generation. Leveraging the high symmetry and even distribution of the Leech lattice, $\Lambda_{24}$-SQ simplifies training and improves reconstruction-compression trade-offs. It outperforms prior art (like BSQ) in image tokenization and compression tasks and extends its benefits to state-of-the-art autoregressive image generation frameworks.