Non-Parametric Quantization
Research code for two visual tokenization methods, BSQ-ViT and its successor L24SQ-ViT, that use spherical vector quantization instead of a learned codebook to tokenize images and video.
Non-Parametric Quantization is a research codebase covering a series of visual tokenizers that quantize onto a sphere instead of using a trained, parametric codebook. BSQ-ViT (Binary Spherical Quantization) tokenizes images and video, and its successor, Spherical Leech Quantization, extends the approach for visual tokenization and generation.
The repository reports image reconstruction results on ImageNet-1K at 256x256, comparing PSNR, SSIM, LPIPS, and reconstruction FID against a reference VAE across different bit budgets, video reconstruction results on UCF-101 against two other named tokenizers, and image synthesis results on ImageNet-1K at 128x128, along with pretrained configs, checkpoints, and their checksums. It installs through a Miniforge3/mamba environment defined in a provided environment file, and links to the arXiv papers, citations, and an image-generation demo built on top of the tokenizer.