CLR-Wire: Towards Continuous Latent Representations for 3D Curve Wireframe Generation (ACM SIGGRAPH 2025)
This repository is the official repository of the paper, CLR-Wire: Towards Continuous Latent Representations for 3D Curve Wireframe Generation.
Xueqi Ma, Yilin Liu, Tianlong Gao, Qirui Huang, Hui Huang,
VCC, CSSE, Shenzhen University
CLR-Wire.mp4
The code is tested in docker enviroment nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04. The following are instructions for setting up the environment in a Linux system from scratch.
First, clone this repository:
git clone git@github.com:qixuema/CLR-Wire.git
Then, create a mamba environment with the yaml file. (Sometimes the conda is a bit slow to solve the dependencies, so mamba is recommended). You could also just use conda as well.
mamba env create -f environment.yml
mamba activate wire
We use the ABC dataset and process its shapes into curve wireframe (aka. curve network). The processed data is stored on huggingface.
To train the model, please use the following commands:
# Train CurveVAE
accelerate launch --config_file src/configs/default_config.yaml train_curve_vae.py --config src/configs/train_curve_vae.yaml
# Train WireframeVAE
accelerate launch --config_file src/configs/default_config.yaml train_wireframe_vae.py --config src/configs/train_wireframe_vae.yaml
# Train Flow Matching
accelerate launch --config_file src/configs/default_config.yaml train_flow_matching.py --config src/configs/train_flow_matching.yaml
To sample the wireframe, download the checkpoint from google drive, extract it to the project root, then run:
python sample.py
If you find our work useful for your research, please consider citing the following papers :)
@inproceedings{CLRWire25,
title = {CLR-Wire: Towards Continuous Latent Representations for 3D Curve Wireframe Generation},
author = {Xueqi Ma and Yilin Liu and Tianlong Gao and Qirui Huang and Hui Huang},
booktitle = {ACM SIGGRAPH},
pages = {},
year = {2025},
}This repo is currently maintained by Xueqi Ma (@qixuema) and is for academic research use only. Discussions and questions are welcome via qixuemaa@gmail.com. Layout and styling adapted from 3dlg-hcvc/omages.