DiGARR is a novel neural rendering framework for robust radiance fields that implements disentangled generation and aggregation methods.

Note: This code is currently under organization and not ready to run directly. Please wait for code organization to complete or contact the authors for the full version.

1. Clone the repository

2. Install dependencies:

pip install -r requirements.txt

The project supports LLFF dataset. Please download the corresponding data and place it in the correct location.

The project provides a convenient training script scripts/train/static_train_hybrid.sh for LLFF dataset training:

# Syntax: ./scripts/train/static_train_hybrid.sh <GPU_ID> <CONFIG_NAME>
# Example: train fern scene
./scripts/train/static_train_hybrid.sh 0 fern

# Train flower scene
./scripts/train/static_train_hybrid.sh 0 flower

# Train horns scene
./scripts/train/static_train_hybrid.sh 0 horns

• GPU_ID: CUDA device ID (e.g., 0, 1, 2...)
• CONFIG_NAME: Configuration file name corresponding to LLFF dataset scene name

• fern - Fern scene
• flower - Flower scene
• fortress - Fortress scene
• horns - Horns scene
• leaves - Leaves scene
• orchids - Orchids scene
• room - Room scene
• trex - T-Rex scene

• Code init
• Code debug, organization and refactoring
• NeRF Blender dataset/training
• Dependency file organization (requirements.txt)
• Pre-trained model preparation

@InProceedings{digarr,
author="Shen, Shihe and Gao, Huachen and Xu, Wangze and Peng, Rui and Tang, Luyang and Xiong, Kaiqiang and Jiao, Jianbo and Wang, Ronggang",
title="Disentangled Generation and Aggregation for Robust Radiance Fields",
booktitle="Computer Vision -- ECCV 2024",
year="2025",
}