Environmental Volumetric Neural Shading of Clouds for Real-Time Rendering

Rikard Olajos
Lund University

Michael Doggett
Lund University

Prashant Goswami
Bleking Institute of Technology

High-Performance Graphics, July 2026

Abstract

We present a high-performance method for real-time relighting of high-fidelity volumetric clouds. Building on Relightable Neural Assets, we introduce key adaptations that enable neural shading of volumetric cloud phenomena within a rasterization-based pipeline. In particular, we incorporate a per-pixel thickness parameter to capture view-dependent opacity and replace the generalizing single light source with a sky illumination model, allowing the network to learn complex atmospheric scattering effects. To achieve real-time performance, we depart from density-field-based volumetric rendering and instead operate on mesh representations combined with a triplane feature encoding. This enables a fully rasterization-driven solution that reproduces volumetric appearance without requiring ray marching or volume integration. We further describe a complete pipeline for converting volumetric cloud assets into a neural representation trained from path-traced supervision. We evaluate our method through an ablation study analyzing both image quality and runtime performance. Our adaptations improve reconstruction quality from 18.13 dB to 22.32 dB while achieving rendering times as low as 3.7 ms per frame. These results demonstrate that our approach enables high-quality, relightable cloud rendering suitable for real-time and performance-critical applications.

Downloads

Paper
Code
Resources
BibTeX entry

@article{olajos-2026-env-volumetric-clouds,
    author = {Olajos, Rikard and Doggett, Michael and Goswami, Prashant},
    title = {Environmental Volumetric Neural Shading of Clouds for Real-Time Rendering},
    year = {2026},
    issue_date = {July 2026},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    volume = {9},
    number = {4},
    url = {https://doi.org/10.1145/3820020},
    doi = {10.1145/3820020},
    abstract = {We present a high-performance method for real-time relighting of high-fidelity volumetric clouds. Building on Relightable Neural Assets, we introduce key adaptations that enable neural shading of volumetric cloud phenomena within a rasterization-based pipeline. In particular, we incorporate a per-pixel thickness parameter to capture view-dependent opacity and replace the generalizing single light source with a sky illumination model, allowing the network to learn complex atmospheric scattering effects. To achieve real-time performance, we depart from density-field-based volumetric rendering and instead operate on mesh representations combined with a triplane feature encoding. This enables a fully rasterization-driven solution that reproduces volumetric appearance without requiring ray marching or volume integration. We further describe a complete pipeline for converting volumetric cloud assets into a neural representation trained from path-traced supervision. We evaluate our method through an ablation study analyzing both image quality and runtime performance. Our adaptations improve reconstruction quality from 18.13 dB to 22.32 dB while achieving rendering times as low as 3.7 ms per frame. These results demonstrate that our approach enables high-quality, relightable cloud rendering suitable for real-time and performance-critical applications.},
    journal = {Proc. ACM Comput. Graph. Interact. Tech.},
    month = jun,
    articleno = {59},
    numpages = {17}
}