Reproduce “Neural Super-Resolution for Real-time Rendering with Radiance Demodulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jia Li, Ziling Chen, Xiaolong Wu, Lu Wang, Beibei Wang, Lei Zhang
NSRD, a new neural super-resolution method for real-time rendering, improves visual quality, temporal stability, and efficiency. It achieves this by separating scene radiance into material and lighting components for targeted super-resolution, robustly handling motion with an occlusion-aware mask, and employing a lightweight recurrent neural network.
Key insights
- Separating radiance into material and lighting components allows for efficient high-resolution texture recovery by super-resolving only the smoother lighting and re-modulating with pre-rendered HR material.
- Explicitly identifying and adaptively handling motion-unreliable regions using an occlusion-aware mask is critical for preventing ghosting artifacts in dynamic scenes.
- A lightweight, frame-recurrent architecture coupled with a tailored temporal loss function can effectively aggregate past frame information for stability without sacrificing real-time performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.