Neural Super-Resolution for Real-time Rendering with Radiance Demodulation

Rendering high-resolution images in real-time applications (e.g., video games, virtual reality) is time-consuming, thus super-resolution technology becomes more and more crucial in real-time rendering. However, it is still challenging to preserve sharp texture details, keep the temporal stability and avoid the ghosting artifacts in the real-time rendering super-resolution. To this end, we introduce radiance demodulation into real-time rendering super-resolution, separating the rendered image or radiance into a lighting component and a material component, due to the fact that the light component tends to be smoother than the rendered image and the high-resolution material component with detailed textures can be easily obtained. Therefore, we perform the super-resolution only on the lighting component and re-modulate with the high-resolution material component to obtain the final super-resolution image. In this way, the texture details can be preserved much better. Then, we propose a reliable warping module by explicitly pointing out the unreliable occluded regions with a motion mask to remove the ghosting artifacts. We further enhance the temporal stability by designing a frame-recurrent neural network to aggregate the previous and current frames, which better captures the spatial-temporal correlation between reconstructed frames. As a result, our method is able to produce temporally stable results in real-time rendering with high-quality details, even in the highly challenging 4 $\times$ 4 super-resolution scenarios.