|
|
--- |
|
|
language: |
|
|
- en |
|
|
license: mit |
|
|
pipeline_tag: text-to-image |
|
|
--- |
|
|
|
|
|
# InfinityCC: Spherical Leech Quantization for Visual Tokenization and Generation |
|
|
|
|
|
This repository hosts **InfinityCC**, a working example showcasing the power of [Non-Parametric Quantization (NPQ)](https://cs.stanford.edu/~yzz/npq/) for ImageNet-1k class-conditioned image generation. |
|
|
|
|
|
The model is based on the paper: [**Spherical Leech Quantization for Visual Tokenization and Generation**](https://huggingface.co/papers/2512.14697) |
|
|
Yue Zhao, Hanwen Jiang, Zhenlin Xu, Chutong Yang, Ehsan Adeli, Philipp Krähenbühl. |
|
|
|
|
|
Project Page: [https://cs.stanford.edu/~yzz/npq/](https://cs.stanford.edu/~yzz/npq/) |
|
|
Code: [https://github.com/zhaoyue-zephyrus/InfinityCC](https://github.com/zhaoyue-zephyrus/InfinityCC) |
|
|
|
|
|
<img src="https://github.com/zhaoyue-zephyrus/InfinityCC/raw/main/assets/npq.png" width="640"> |
|
|
|
|
|
## Introduction |
|
|
In this work, we explore Spherical Leech Quantization ($\Lambda_{24}$-SQ), a non-parametric quantization method rooted in lattice coding. This approach simplifies the training recipe and improves the reconstruction-compression tradeoff, thanks to its high symmetry and even distribution on the hypersphere. It has demonstrated better reconstruction quality than prior art in image tokenization and compression tasks, with improvements extending to state-of-the-art auto-regressive image generation frameworks. InfinityCC serves as a practical demonstration of this powerful quantization technique for visual generation. |
|
|
|
|
|
## Installation |
|
|
|
|
|
We use [uv](https://docs.astral.sh/uv/) to manage all dependencies. |
|
|
|
|
|
```bash |
|
|
uv sync |
|
|
source .venv/bin/activate |
|
|
``` |
|
|
|
|
|
To evaluate ImageNet using the ADM evaluator, run the following command lines: |
|
|
```bash |
|
|
mkdir third_party/ && cd third_party/ |
|
|
git clone https://${GIT_TOKEN}@github.com/openai/guided-diffusion.git |
|
|
cd guided-diffusion/evaluations |
|
|
wget https://openaipublic.blob.core.windows.net/diffusion/jul-2021/ref_batches/imagenet/256/VIRTUAL_imagenet256_labeled.npz |
|
|
``` |
|
|
|
|
|
## Results |
|
|
|
|
|
### InfinityCC Performance |
|
|
| model | Resolution | #layers | Tokenizer (HF weights🤗) | VAR Model (HF weights🤗) | FID | |
|
|
|:----------:|:-----:|:--------:|:---------:|:-----------------------------------------------------------------------------------:|:----:| |
|
|
| InfinityCC | 256 | 12 | [bitvae_l24_xl](https://huggingface.co/zhaoyue-zephyrus/InfinityCC_L24SQ/tree/main/tokenization/infinity_l24_stage1_xl) | [infinitycc_12layer_weights](https://huggingface.co/zhaoyue-zephyrus/InfinityCC_L24SQ/tree/main/generation/infinitycc_12layer_256x256_l24_xl_ep50_cce_zloss_improved_schedule_dion) | 6.66 | |
|
|
| InfinityCC | 256 | 24 | [bitvae_l24_xl_vf](https://huggingface.co/zhaoyue-zephyrus/InfinityCC_L24SQ/tree/main/tokenization/infinity_l24_stage1_xl_vf) | [infinitycc_24layer_weights](https://huggingface.co/zhaoyue-zephyrus/InfinityCC_L24SQ/tree/main/generation/infinitycc_24layer_256x256_l24_xl_vf_ep350_cce_zloss_improved_schedule_dion_unsharedaln) | 2.21 | |
|
|
| InfinityCC-2B | 256 | 32 | [TBD]() | [TBD]() | 1.80 | |
|
|
|
|
|
## Citation |
|
|
If our work assists your research, feel free to give us a star ⭐ or cite us using: |
|
|
|
|
|
```bibtex |
|
|
@article{zhao2025spherical, |
|
|
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 Krähenbühl, Philipp}, |
|
|
journal={arXiv preprint arXiv:2512.14697}, |
|
|
year={2025} |
|
|
} |
|
|
``` |
|
|
|
|
|
## License |
|
|
This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details. |
