Neural Global Illumination: Interactive Indirect Illumination Prediction Under Dynamic Area Lights

Area of Science:

• Computer Graphics
• Artificial Intelligence

Background:

• Global illumination rendering is computationally intensive.
• Simulating complex lighting effects like interreflection and caustics requires sophisticated techniques.

Purpose of the Study:

• To develop a novel method for fast rendering of full global illumination in static scenes.
• To enable dynamic viewpoint and area lighting with high-fidelity global illumination effects.

Main Methods:

• Utilizing a deep rendering network to model the mapping from shading points to global illumination.
• Employing a neural-network-friendly input representation with positional encoding for high-frequency effects.
• Integrating a screen-space neural buffer for efficient global information sharing.

Main Results:

• Demonstrated fast rendering of complex global illumination effects.
• Successfully rendered scenes with multiple-bounce glossy interreflection, color bleeding, and caustics.
• Achieved high-quality fitting with a compact neural network.

Conclusions:

• Neural global illumination offers an efficient and effective solution for rendering complex lighting.
• The proposed method advances the state-of-the-art in real-time global illumination rendering.
• This technique is applicable to a wide variety of scenes with intricate lighting phenomena.