GCRayDiffusion

Li-Heng Chen, Zi-Xin Zou, Chang Liu, Tianjiao Jing, Yan-Pei Cao, Shi-Sheng Huang†, Hongbo Fu, Hua Huang

† Corresponding Author

This repository contains the source code of GCRayDiffusion:Pose-Free Surface Reconstruction via Geometric Consistent Ray Diffusion (ICCV 2025)

Contents (Coming soon)

• Training Code
• Installation Instructions
• End-to-End Training Demo
• Coarse and Fine Tuning Instruction

🔨 Installation

Clone the repository:

git clone https://github.com/CountNemoChan/GCRayDiffusion.git
cd GCRayDiffusion

Create a conda environment(optional):

conda create -n gcraydiffusion python=3.9
conda activate gcraydiffusion

Install dependencies:

# pytorch (select correct CUDA version)
pip install torch torchvision --index-url https://download.pytorch.org/whl/{your-cuda-version}

# other dependencies
pip install -r requirements.txt

💡 Training Steps

Step 1: Prepare your training dataset

Our training data comes from the open-source dataset Objaverse. Due to the access limitation, I will show you some sample data. You may need to modify the Dataloader later.

Step 2: Start training

python launch.py --config configs/training_config.yaml --train --gpu 0

Complete details of all training parameters are provided in the Config_File.

Training Demo

To facilitate training the model on the complete dataset, we provide a training demo to help clarify the training process.

Just run:

python launch.py --config configs/sample_training_config.yaml --train --gpu 0