需先安装 Miniforge3 才能用 mamba 创建环境;预训练权重需从 Google Drive 手动下载;部分视频与合成任务的 config/ckpt 仍标注为 TBA 未放出

npq-vit
zhaoyue-zephyrusMIT活跃维护面向视觉 tokenizer 的球面量化方案,包含二值球面量化和 Leech 球面量化两种方法,用于压缩视觉 token 并提升表示效率。
[ICLR 2025] Binary Spherical Quantization + [CVPR 2026] Leech Spherical Quantization
mamba env create -f bsqvit-env.yaml
mamba activate bsqvitREADME 未说明
用二值球面量化和 Leech 球面量化做图像视频 tokenizer 的研究代码库
面向视觉 tokenizer 的球面量化方案,包含二值球面量化和 Leech 球面量化两种方法,用于压缩视觉 token 并提升表示效率。
核心能力
- 用二值球面量化(BSQ)做图像与视频 tokenization
- 提出 Λ24 球面 Leech 量化(L24SQ)用于视觉 token 化与生成
- 提供 IN-1K 图像重建、UCF-101 视频重建的预训练权重
- 支持 18/36 bit 等不同码率的重建配置
- 附图像合成与视频压缩的评测结果
容易踩的坑
暂无 release 数据(可能没有发版,或尚未同步)。
用户点评
登录用户的真实使用记录,可按版本和硬件筛选还没有人点评。第一个说出真实体验的人最有价值。
Non-Parametric Quantization
I wrote a blog about the motivation behind this series of work.
Λ24-SQ-ViT
Spherical Leech Quantization for Visual Tokenization and Generation
Yue Zhao1,2, Hanwen Jiang1,3, Zhenlin Xu4, Chutong Yang1, Ehsan Adeli2, Philipp Krähenbühl1
1UT Austin, 2Stanford University, 3Adobe Research, 4Mistral AI
arxiv | bibtex | Image Generation with Infinity+L24SQ
🍰 BSQ-ViT
You can pronouce BSQ-ViT like "biskvit" (a kind of Russian sponge cake) or simply "biscuit".
Image and Video Tokenization with Binary Spherical Quantization
Yue Zhao1, Yuanjun Xiong2, Philipp Krähenbühl1
1UT Austin, 2Predera
arxiv | bibtex
Installation
-
Install Miniforge3
-
Create the environment
mamba env create -f bsqvit-env.yaml
mamba activate bsqvit
Main Results
Image Reconstruction (IN-1K val 256x256)
| Use approx. (Eq 8) | #bits | PSNR↑ | SSIM↑ | LPIPS↓ | rFID↓ | config & ckpt | md5sum | |
|---|---|---|---|---|---|---|---|---|
| SDXL-VAE | N/A | 64 | 25.38 | .7276 | .0666 | 0.72 | External | N/A |
| BSQ-ViT | 18 | 24.79 | .7319 | .0836 | 1.34 | Gdrive | 7abf5a | |
| BSQ-ViT (EMA) | 18 | 24.80 | .7314 | .0820 | 1.23 | Gdrive | 7abf5a | |
| BSQ-ViT | ✓ | 18 | 25.36 | .7578 | .0761 | 1.14 | Gdrive | 8f5422 |
| BSQ-ViT (EMA) | ✓ | 18 | 25.80 | .7680 | .0729 | 1.30 | Gdrive | 8f5422 |
| BSQ-ViT | ✓ | 36 | 27.88 | .8410 | .0432 | 0.41 | Gdrive | b5ce5f |
| BSQ-ViT (EMA) | ✓ | 36 | 28.14 | .8448 | .0400 | 0.45 | Gdrive | b5ce5f |
| L24SQ-ViT (EMA) | N/A | ~17.6 | Gdrive | 73ab9e |
Video Reconstruction (UCF-101 16x128x128)
| #bits | PSNR↑ | SSIM↑ | LPIPS↓ | rFVD↓ | config & ckpt | |
|---|---|---|---|---|---|---|
| MAGVIT-L | 10 | 22.0 | .7010 | .0990 | 25 | N/A |
| MAGVITv2 | 18 | - | - | .0694 | 16.12 | N/A |
| MAGVITv2 (deeper) | 18 | - | - | .0537 | 8.62 | N/A |
| BSQ-bcViT | 18 | 32.08 | .9421 | .0244 | 8.08 | TBA |
| BSQ-bcViT | 36 | 33.80 | .9606 | .0159 | 4.10 | TBA |
Image Synthesis (IN-1K 128x128)
| FID↓ | IS↑ | Prec↑ | Rec↑ | pre-computed samples | config & ckpt | |
|---|---|---|---|---|---|---|
| BigGAN | 6.02 | 145.8 | 0.86 | 0.35 | External | External |
| ADM | 5.91 | 93.3 | 0.70 | 0.65 | External | External |
| Ours BSQ-ViT + Masked-LM | 5.44 | 139.6 | 0.80 | 0.50 | UTBox | TBA |
Video Compression (MCL-JCV 360P)
License
Citing NPQ-ViT
@article{zhao2024lsqvit,
title={Spherical Leech Quantization for Visual Tokenization and Generation},
author={Zhao, Yue and Jiang, Hanwen and Xu, Zhenlin and Yang, Chutong and Adeli, Ehsan and Kr{\"a}henb{\"u}hl, Philipp},
journal={arXiv preprint arXiv:2512.14697},
year={2025}
}
@article{zhao2024bsqvit,
title={Image and Video Tokenization with Binary Spherical Quantization},
author={Zhao, Yue and Xiong, Yuanjun, and Kr{\"a}henb{\"u}hl, Philipp},
journal={arXiv preprint arXiv:2406.07548},
year={2024}
}