Neural BRDF Importance Sampling by Reparameterization

This repo contains the implementation of SIGGRAPH 2025 paper: Neural BRDF Importance Sampling by Reparameterization.

Paper | Citation

Setup

• CUDA 11.7
• pytorch 2.3.1
• lightning 2.1.3
• mitsuba 3.4.0

Pre-trained models

The pre-trained weights for both RGL and NeuSample dataset can be found in here

Usage

1. Edit configs/rgl.yaml or configs/neusample.yaml to setup data path and configure a training.
2. To train the reparameterization model, run:

python train.py --experiment_name <experiment-name> --configs <config-file> --device <gpu-id> --max_epochs <number-of-epochs>

3. To train the pdf approximation, run:

python train_mis.py --experiment_name <experiment-name> --configs <config-file> --device <gpu-id> --max_epochs <number-of-epochs>

4. demo/demo.ipynb contains a rendering example using mitsuba 3.

Citation

@inproceedings{wu2025neural,
      title={Neural BRDF Importance Sampling by Reparameterization},
      booktitle = {SIGGRAPH},
      author={Liwen Wu and Sai Bi and Zexiang Xu and Hao Tan and Kai Zhang and Fujun Luan and Haolin Lu and Ravi Ramamoorthi},
      year={2025}
}