pipeline_tag: text-to-image
SphereAR: Hyperspherical Latents Improve Continuous-Token Autoregressive Generation
This repository contains the official PyTorch implementation of the paper Hyperspherical Latents Improve Continuous-Token Autoregressive Generation.
SphereAR proposes a simple yet effective approach to continuous-token autoregressive (AR) image generation. It addresses issues like heterogeneous variance in VAE latents, which is amplified during AR decoding, by constraining all AR inputs and outputs---including after Classifier-Free Guidance (CFG)---to lie on a fixed-radius hypersphere (constant $\ell_2$ norm) via hyperspherical VAEs. This approach removes the scale component, thereby stabilizing AR decoding.
The model is a pure next-token AR generator with raster order. Empirically, on ImageNet 256×256 generation, SphereAR-H (943M) achieves a new state-of-the-art for AR models, reaching FID 1.34.
For more details, including implementation, training, and evaluation scripts, please refer to the official GitHub repository.
Model Checkpoints
Pre-trained model checkpoints are available:
| Name | params | FID (256x256) | weight |
|---|---|---|---|
| S-VAE | 75M | - | vae.pt |
| SphereAR-B | 208M | 1.92 | SphereAR_B.pt |
| SphereAR-L | 479M | 1.54 | SphereAR_L.pt |
| SphereAR-H | 943M | 1.34 | SphereAR_H.pt |
Citation
If you find this work useful, please consider citing the paper:
@article{ke2025hyperspherical,
title={Hyperspherical Latents Improve Continuous-Token Autoregressive Generation},
author={Guolin Ke and Hui Xue},
journal={arXiv preprint arXiv:2509.24335},
year={2025}
}