NEnv: Neural Environment Maps for Global Illumination

Official repository of "NEnv: Neural Environment Maps for Global Illumination"

Carlos Rodriguez-Pardo*, Javier Fabre*, Elena Garces, Jorge Lopez-Moreno

Computer Graphics Forum (Proceedings of the Eurographics Symposium on Rendering), June 2023

Project Website

Paper Link

Dataset and Interactive Website

We introduce NEnv, an invertible and fully differentiable neural method which achieves high-quality reconstructions for environment maps and their probability distributions. NEnv is up to two orders of magnitude faster to sample from than analytical alternatives, providing fast and accurate lighting representations for global illumination using Multiple Importance Sampling. Our models can accurately represent both indoor and outdoor illumination, achieving higher generality than previous work on environment map approximations.

Requirements

Please use pip to install the required packages. pip install -r requirements.txt

Usage

To evaluate or sample from a pre-trained normalizing flow, please see an example in NEnv/Scripts/eval_flow.py or NEnv/Scripts/eval_compression.py. Just change the path to your desired pre-trained flow.

To train a flow from an input environment map, please follow NEnv/Scripts/train_nenv.py.

To train a compression_model from an input environment map, please follow NEnv/Scripts/train_nenv_compression.py.

Dataset

Please visit the official website to find the dataset of pre-trained models.

Coming Soon

In planned release order:

• Pre-processing algorithms
• PyTorch3D integration
• PyPI package

Citation

Please cite our publication if you end up using any of this code in your research.

@inproceedings{Rodriguez-Pardo_2023_EGSR,
author = {Rodriguez-Pardo, Carlos and Fabre, Javier and Garces, Elena and Lopez-Moreno, Jorge},
title = {NEnv: Neural Environment Maps for Global Illumination},
booktitle = {Computer Graphics Forum (Eurographics Symposium on Rendering Conference Proceedings)},
year = {2023},
publisher = {The Eurographics Association and John Wiley & Sons Ltd.},
ISSN = {1467-8659},
DOI = {10.1111/cgf.14883}
}

Acknowledgements

Our implementation is based on Neural Spline Flows.