Taming High-Resolution Auxiliary G-Buffers for Deep Supersampling of Rendered Content.

Area of Science:

• Computer Graphics
• Image Processing
• Artificial Intelligence

Background:

• High-resolution rendering faces computational challenges, often addressed by rendering at lower resolutions and upsampling.
• Existing supersampling techniques underutilize high-frequency information present in high-resolution G-buffers.

Purpose of the Study:

• To investigate leveraging high-resolution G-buffer information for improved supersampling visual quality.
• To develop a neural network-based real-time supersampling method that maximizes detail recovery.

Main Methods:

• Proposed a neural network incorporating gated G-buffers encoder, G-buffers attended encoder, and reflection-aware loss.
• Developed an occlusion-aware blender to improve temporal stability by rectifying dis-occluded features.
• Utilized high-frequency information from high-resolution G-buffers for faithful detail recovery from aliased inputs.

Main Results:

• The proposed method significantly enhances visual fidelity in high-resolution reconstructions compared to state-of-the-art methods.
• Achieved superior results even for challenging 4x4 upsampling tasks.
• Demonstrated compute efficiency alongside improved image quality.

Conclusions:

• The developed neural network effectively harnesses high-resolution G-buffer information for superior real-time supersampling.
• The method offers a significant advancement in recovering high-frequency details and improving temporal stability.
• This approach provides a compute-efficient solution for high-quality real-time rendering.