The official implement of "MLI-NeRF: Multi-Light Intrinsic-Aware Neural Radiance Fields".

Yixiong Yang, Shilin Hu, Haoyu Wu, Ramon Baldrich, Dimitris Samaras, Maria Vanrell

International Conference on 3D Vision 2025

| Paper |

https://github.com/user-attachments/assets/0ae8436f-a2af-4c07-8a96-b504db106844

Requirements

We use Neuralangelo as the baseline model. Therefore, we recommend following the installation instructions on the Neuralangelo website (without the need to install COLMAP).

Tips:

1. The code requires tiny-cuda-nn, which needs the CUDA version of PyTorch to be consistent with the CUDA version installed on your system.
2. Use pip to install pymcubes==0.1.4.

Run on the synthetic dataset

One example from the synthetic dataset. (The datasets can be downloaded from this link) Modify the path in CONF_a and CONF_b.

Inference with Trained Model

The trained model can be downloaded from this link. A video can be generated with the following command:

CONF_b="syn_hotdog_b"
python test.py --config=projects/NeuralLumen/configs/${CONF_b}.yaml --show_pbar --single_gpu --inference_mode video_train_0_67

Train and test

./run_synthetic.sh

Run on the real object dataset

Let's use one example from the real object dataset.

Download dataset from: https://github.com/iamNCJ/NRHints Modify the path in CONF_a and CONF_b.

Inference with Trained Model

The trained model can be downloaded from the previous link.

CONF_b="NRHints_Pikachu_b"
python test.py --config=projects/NeuralLumen/configs/${CONF_b}.yaml --show_pbar --single_gpu --inference_mode video_train_0_67

Train and test

./run_real.sh

Run on the ReNe dataset

The ReNe dataset can be downloaded from ReNe.
We provide a script, projects/NeuralLumen/scripts/convert_rene_direct_to_json.py, to generate the JSON files.
Note: This script requires the ReNe dataset package to be installed beforehand.

The generated JSON files are also included in dataset_rene.

Inference with Trained Model

The trained model can be downloaded from the previous link.

CONF_b="rene_savannah_b"
python test.py --config=projects/NeuralLumen/configs/${CONF_b}.yaml --show_pbar --single_gpu --inference_mode video_train_0_67

Train and test

./run_rene.sh

Citation

Please cite this repository as follows if you find it helpful for your project :):

@inproceedings{yang2024mlinerf,
  title={MLI-NeRF: Multi-Light Intrinsic-Aware Neural Radiance Fields},
  author={Yixiong Yang and Shilin Hu and Haoyu Wu and Ramon Baldrich and Dimitris Samaras and Maria Vanrell},
  booktitle={The International Conference on 3D Vision},
  year={2025}
}

Acknowledgments

Some codes are borrowed from Neuralangelo and NRHints. Thanks for their great works.