AESOP — Human–Camera Generation

In brief

AESOP uses an asymmetric Human-first architecture to unify Camera generation from given Human motion and joint Human–Camera generation, while preserving Human motion and enabling explicit control over Camera translation intensity.

AESOP generates camera trajectories given human motion, jointly generates humans and cameras, and controls camera translation intensity.

Demo

Method

An independent human generator and a shared, human-conditioned camera generator.

AESOP pipeline showing asymmetric human and camera representations, independent human flow, and shared camera flow with translation-intensity conditioning.

Comparisons

Intensity Control

Vary camera travel with intensity a. Human motion, camera text, and noise are shared.

External Examples

Trained on PulpMotion. Applied to HumanML3D and HY-Motion.

BibTeX

@article{lin2026aesop,
  title={AESOP: Asymmetric Human-Camera Generation with Translation-Intensity Control},
  author={Lin, Jingzhong and Wang, Zhanke and Li, Heng and Liu, Wenxiang
    and Zhang, Zhao and Tang, Kecheng and Xiang, Dongdong and Wang, Changbo
    and Kang, Di and Guo, Chunchao and Bao, Linchao and He, Gaoqi},
  journal={arXiv preprint arXiv:2609.37229},
  year={2026},
  url={https://arxiv.org/abs/2609.37229}
}