TL;DR
Neural Cone Radiosity (NCR) introduces a highly efficient method for real-time global illumination on glossy surfaces by replacing traditional point-based ray queries with scale-aware ray cone encoding. By combining a pre-filtered multi-resolution hash grid with a smart clustering algorithm, NCR delivers noise-free, temporally stable glossy reflections at interactive frame rates.
What Problem Does This Paper Address?
Achieving realistic, real-time global illumination (GI) remains one of the most challenging problems in computer graphics. While recent Neural Global Illumination (NGI) and Neural Radiosity (NR) methods have made massive strides, glossy materials—like brushed metals, frosted glass, or varnished wood—remain a notorious failure case.
Glossy surfaces reflect light in narrow, view-dependent "lobes." Existing neural methods typically face a harsh trade-off: they either trace secondary rays and query a neural radiance cache (which introduces massive Monte Carlo noise) or evaluate the network directly at the primary hit point (which smooths out and over-blurs high-frequency reflections). NCR tackles this noise-bias dilemma to render sharp, physically consistent glossy inter-reflections in real time.
Methods
NCR upgrades point-based neural queries to a scale-aware spatial integration framework through three core components:
- Cone Encoding for Glossy Surfaces: Instead of treating a ray-surface intersection as an infinitesimal point, NCR traces a ray cone whose aperture is dynamically determined by the material's surface roughness. It queries a multi-resolution hash grid using both the footprint's center and its spatial extent, effectively pre-filtering the incident radiance over the BSDF lobe.
- Clustering Approximation: Because real-world geometry is complex, a cone's intersection with a scene rarely forms a perfect circle. To handle this efficiently, NCR traces a small set of stochastic rays and groups them using a fast 1D K-Means clustering algorithm based on marching distances. The radiance from these representative clusters is then weighted and averaged.
- Dual-Branch Network Architecture: The system splits the workload into two specialized branches: a diffuse branch that captures low-frequency spatial lighting, and a compact glossy branch utilizing the cone encoding. A lightweight modulation network ($\mathcal{F}_{mod}$) dynamically blends their outputs based on surface roughness and reflectance.
Datasets
The authors evaluated their method on several modified scenes from the standard Bitterli dataset:
- Bathroom: Features a frosted mirror, shiny tiled floor, and a ceramic bathtub.
- Cornell-box: Includes a frosted mirror reflecting plastic bunny and dragon models alongside complex wall textures.
- Kitchen: Contains a brushed metal extractor hood and various glossy ceramic geometries.
- Living-room: Showcases varnished wood surfaces and a polished mirror.
- Veach-ajar: Comprises a glossy floor and pots with sharp specular highlights.
Key Insights and Ideas
- Glossy Reflection as Convolution: The authors treat glossy reflection as a mathematical convolution of incident radiance and the BSDF lobe. By offloading this convolution to a pre-filtered multi-resolution hash grid, the neural network doesn't have to struggle to fit complex, high-frequency angular distributions.
- 1D vs. 3D Clustering: Clustering ray intersections in 3D space is computationally expensive for real-time pipelines. NCR elegantly bypasses this by performing 1D K-Means clustering strictly on the marching distances of the rays, drastically cutting down overhead while successfully grouping depth variations.
- Network Downsizing: By delegating the complex directional modeling to the pre-filtered glossy branch, the authors were able to shrink the primary MLP network size. This architectural efficiency keeps the runtime footprint small.
Key Results
- Superior Visual Fidelity: NCR consistently outperforms vanilla Neural Radiosity and standard path tracing in image quality. For instance, in the Cornell-box scene, NCR achieved a Mean Absolute Percentage Error (MAPE) of 0.057, compared to NR's 0.073.
- Temporal Stability: While path tracing combined with state-of-the-art denoisers (like Intel's OIDN) can produce comparable static image quality, denoisers often introduce distracting flickering artifacts during interactive camera movement. NCR delivers superior temporal stability with completely noise-free animations.
- Interactive Performance: NCR runs at interactive frame rates (ranging from ~37ms to 116ms per frame on an NVIDIA RTX 3090 depending on scene complexity), making it highly competitive with vanilla NR despite the advanced glossy transport simulation.
Limitations and Future Work
- Per-Scene Training: Like many coordinate-based neural representations, the model must be trained from scratch for each specific scene configuration, lacking instant zero-shot generalization across different environments.
- Glossy Dielectric Caustics: NCR struggles to accurately reproduce complex caustic effects on glossy dielectric (refractive) materials due to the inherent difficulty of sampling light sources effectively after multi-bounce light transport.
- Future Directions: The authors plan to explore integrating their cone encoding technique with generalizable, scene-agnostic neural rendering frameworks to make the method viable for industrial-scale, plug-and-play applications.
Summary updated