This repository is the official implementation of ETCH, a novel body fitting pipeline that estimates cloth-to-body surface mapping through locally approximate SE(3) equivariance, encoding tightness as displacement vectors from the cloth to the underlying body.
News 🚩
[2025-08-04] We release the All-in-One model, which is trained on the 4D-Dress dataset, CAPE dataset, and Generative dataset, totally 94501 samples. Please download the all-in-one model from here.
[2025-08-04] We release the code for ETCH, please feel free to have a try!
Overview
Our key novelty is modeling cloth-to-body SE(3)-equivariant tightness vectors for clothed humans, abbreviated as ETCH, which resembles ``etching'' from the outer clothing down to the inner body.
Following this outer-to-inner mapping, ETCH regresses sparse body markers, simplifying clothed human fitting into an inner-body marker fitting task.
please note that we placed data samples in the datafolder folder for convenience.
Generate Anchor Points with Tightness Vectors (for training)
python scripts/generate_infopoints.py
Get splitted ids (pkl file)
python scripts/get_splitted_ids_{datasetname}.py
For body_models, please download with this link, and place it under the datafolder/ folder.
please note that before the above processes, there are some preprocessing steps on the original data:
for 4D-Dress dataset, we apply zero-translation mesh.apply_translation(-translation) to the original scan and the body model;
for CAPE dataset, we used the processed meshes extracted from PTF, in which we noticed that the SMPL body meshes are marginally different from the original SMPL body meshes but more precise.
please refer to the datafolder folder for more details.
Training 🚀
bash
1CUDA_VISIBLE_DEVICES=0 python src/train.py --batch_size 2 --i datasetname_settingname
2# batch_size should <= num_data, if you just have the sample data, you can set batch_size to 1
Evaluation 📊
bash
1CUDA_VISIBLE_DEVICES=0 python src/eval.py --batch_size 3 --model_path path_to_pretrained_model --i datasetname_settingname
23# please note that the train_ids has no overlap with the val_ids, the sample data is from train_ids, so if you want to test the pretrained model on the sample data, you should set the activated_ids_path to the train_ids.pkl file for successful selection.
Pretrained Model used in the paper
Please download the pretrained model used in the paper from here.
🔥 All-in-One Model 🔥
We provide the All-in-One model, which is trained on the 4D-Dress dataset, CAPE dataset, and Generative dataset, totally 94501 samples. Please download the all-in-one model from here.
For demo inference, you can use the following command:
Please note that during the training of All-in-One model and in the inference_demo.py file, we centering the scan as input, and re-center the predicted SMPL mesh to the original scan.
For more details, please refer to the src/inference_demo.py file.
We also provide the animation function, which can be used to animate the scan with the predicted SMPL mesh. please refer to the src/animation.py file for more details.
Citation
bibtex
1@inproceedings{li2025etch,
2 title = {{ETCH: Generalizing Body Fitting to Clothed Humans via Equivariant Tightness}},
3 author = {Li, Boqian and Feng, Haiwen and Cai, Zeyu and Black, Michael J. and Xiu, Yuliang},
4 booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
5 year = {2025}
6}
Acknowledgments
We thank Marilyn Keller for the help in Blender rendering, Brent Yi for fruitful discussions, Ailing Zeng and Yiyu Zhuang for HuGe100K dataset, Jingyi Wu and Xiaoben Li for their help during rebuttal and building this open-source project, and the members of Endless AI Lab for their help and discussions. This work is funded by the Research Center for Industries of the Future (RCIF) at Westlake University, the Westlake Education Foundation. Yuliang Xiu also received funding from the Max Planck Institute for Intelligent Systems.
Kudos to all of our amazing contributors! This open-source project is made possible by the contributions of the following individuals:
License
Code License: The ETCH source code is released under the MIT License.
Pretrained Models License: The pretrained models provided in this repository are for non-commercial use only. The release of pretrained models follows the same licensing terms as the datasets used for training. Please refer to the licensing terms of the 4D-Dress dataset and CAPE dataset for more details.
Disclosure
While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society.
