MNSS: Neural Supersampling Framework for Real-Time Rendering on Mobile Devices.

Area of Science:

• Computer Graphics
• Artificial Intelligence

Background:

• Neural supersampling enhances image quality but is computationally demanding.
• Existing methods are too resource-intensive for mobile devices and real-time applications.

Purpose of the Study:

• To develop a lightweight neural supersampling framework for real-time rendering on smartphones.
• To enable high-quality image reconstruction from low-resolution input within a real-time budget.

Main Methods:

• Proposed an alternate sub-pixel sample pattern during rasterization for sampling efficiency.
• Implemented a history buffer with a check-and-reuse scheme for temporal stability.
• Created a new dataset and introduced the inter-frame structural similarity (IF-SSIM) metric for temporal stability evaluation.

Main Results:

• The framework reconstructs high-resolution, anti-aliased images from low-resolution input.
• Achieved high image quality with a small reconstruction model.
• Demonstrated superior performance and temporal stability compared to existing solutions.

Conclusions:

• This is the first real-time neural supersampling research for mobile devices.
• The proposed method is efficient and suitable for resource-constrained platforms.
• The framework significantly improves both rendering performance and temporal stability.