- [2025.12.03] 🔥 Paper is available on Arxiv
OmniPerson is a unified framework for identity-preserving pedestrian generation. It aims to generate diverse and realistic pedestrian images while maintaining the identity of the person across different poses and viewpoints.
It supports various tasks such as:
- Pose Transfer: Transfer the pose of a person to another while preserving their identity. (supports Skeleton or Smplx poses inputs)
- Modality Transfer: Transfer the modality of a person to another, such as from RGB to Infrared.
- Multi-references: Support multiple reference images for a more precise generation. (no upper limit on the number of references)
- Video Generation: Generate a sequence of images that form a coherent video clip. (both IR and RGB)
- Others: Support various other tasks such as super-resolution, occlusion handling, and more.
| Milestone | Status |
|---|---|
| Pretrained weights & Inference code | TBD |
| Training code | TBD |
If you find our work useful for your research, please consider citing the paper:
@article{ma2025omniperson,
title={OmniPerson: Unified Identity-Preserving Pedestrian Generation},
author={Ma, Changxiao and Yuan, Chao and Shi, Xincheng and Ma, Yuzhuo and Zhang, Yongfei and Zhou, Longkun and Zhang, Yujia and Li, Shangze and Xu, Yifan},
journal={arXiv preprint arXiv:2512.02554},
year={2025}
}