Interactive Simulation of Scattering Effects in Participating Media Using a Neural Network Model

Area of Science:

• Computer Graphics
• Computational Physics
• Artificial Intelligence

Background:

• Photorealistic image creation relies on accurately rendering participating media, which exhibit light scattering.
• Simulating multiple scattering events in translucent materials with short mean-free-paths (mfp) is computationally demanding.
• Current methods use precomputed tables, offering speed but incurring high storage costs.

Purpose of the Study:

• To develop a novel, memory-efficient representation for simulating double and multiple scattering effects in participating media.
• To reduce the computational intensity and storage requirements of rendering complex light interactions.

Main Methods:

• A neural network model is proposed to encode the scattering response of homogeneous participating media.
• The neural network predicts double and multiple scattering effects at runtime after a preprocessing step.
• The method is implemented on a GPU and demonstrated with Virtual Ray Lights (VRL).

Main Results:

• The neural network encodes scattering effects using only 23.6 KB of memory.
• The algorithm achieves a rendering speed of 50 ms per frame in typical scenes.
• The results closely match reference rendering quality.

Conclusions:

• The proposed neural network approach offers a highly memory-efficient and computationally fast solution for simulating scattering in participating media.
• This method significantly improves upon traditional table-based precomputation techniques.
• The approach enables high-quality, efficient rendering of complex translucent materials.