Inverse Garment Design
with Visual Agents

Nanxiang Jiang1 Yunuo Chen1,5 Zeng Tao1 Zizhang Li2
Qi Guo3 Jiajun Wu2 Yin Yang4,5 Chenfanfu Jiang1,5

Problem Setting

Simulation-ready garment reconstruction from a single image

Given a single in-the-wild image, we recover every garment of the outfit as a sewing pattern. The patterns are simulation-ready, so they drape and animate on the body, and directly serve 3D and 4D generation.

Garments in Motion

Our pipeline is not limited by camera view, body pose, or the number of garment layers. From a single in-the-wild image, it recovers simulation-ready garments that move naturally with the body in 4D scenes.

Friends dance on a sunny beach

A crew rehearses a routine above the city

Flowing robes sway on an autumn lakeside walk

Runners race along a mountain road

Interactive Viewer

Drag to turn the outfit in 3D. Click a layer below to peel the outfit down to it, and the sewing pattern on the left follows.

 

Comparisons

ChatGarment, ISP, and GarmentRecovery do not generate textures, so we show only their geometry.

Layered Outfits

Select Scene:

Diverse Poses and Body Shapes

Select Scene:

Bibtex

@article{jiang2026inversegarmentdesign,
  title={Inverse Garment Design with Visual Agents},
  author={Jiang, Nanxiang and Chen, Yunuo and Tao, Zeng and Li, Zizhang and Guo, Qi and Wu, Jiajun and Yang, Yin and Jiang, Chenfanfu},
  journal={arXiv preprint arXiv:2610.15582},
  year={2026},
  url={https://arxiv.org/abs/2610.15582}
}