In this paper, we propose a novel equivariant image modeling framework that inherently aligns optimization targets across subtasks in autoregressive image modeling by leveraging the translation invariance of natural visual signals. Our method introduces:
Column-wise tokenization which enhances translational symmetry along the horizontal axis.
Autoregressive generative models using windowed causal attention which enforces consistent contextual relationships across positions.
Evaluated on class-conditioned ImageNet generation at 256×256 resolution, our approach achieves performance comparable to state-of-the-art AR models while using fewer computational resources. Moreover, our approach significantly improving zero-shot generalization and enabling ultra-long image synthesis.
Bibtex
bibtex
1@misc{dong2025equivariantimagemodeling,
2 title={Equivariant Image Modeling},
3 author={Ruixiao Dong and Mengde Xu and Zigang Geng and Li Li and Han Hu and Shuyang Gu},
4 year={2025},
5 eprint={2503.18948},
6 archivePrefix={arXiv},
7 primaryClass={cs.CV},
8 url={https://arxiv.org/abs/2503.18948},
9}
