Lightweight Bilateral Convolutional Neural Networks for Interactive Single-Bounce Diffuse Indirect Illumination.

Area of Science:

• Computer Graphics
• Real-time Rendering
• Global Illumination

Background:

• Physically correct global illumination is essential for realistic rendering but computationally expensive.
• Existing methods struggle to achieve interactive performance due to time-consuming ray sampling and complex neural networks.

Purpose of the Study:

• To develop a novel method for generating plausible single-bounce indirect illumination for dynamic scenes at interactive framerates.
• To address the limitations of current approaches in achieving both quality and speed.

Main Methods:

• Compute direct illumination first.
• Employ a lightweight neural network with bilateral convolution layers to predict screen-space indirect illumination.
• Utilize essential inputs: direct illumination, surface normals, and 3D positions.
• Maintain temporal coherence between frames efficiently without heavy recurrent connections.

Main Results:

• The proposed method generates plausible single-bounce indirect illumination for dynamic scenes.
• Achieves higher quality and better temporal coherence compared to state-of-the-art methods.
• Runs at interactive framerates, enabling real-time applications.

Conclusions:

• The novel method effectively generates high-quality, temporally coherent indirect illumination in real-time.
• Lightweight neural networks with specialized layers offer a viable solution for accelerating physically-based rendering.
• This approach significantly advances the state-of-the-art in interactive global illumination for dynamic scenes.