Joint Denoising and Upscaling via Multi-branch and Multi-scale Feature Network

Pawel Kazmierczyk, Sungye Kim, Wojciech Uss, Wojciech Kalinski, Tomasz Galaj, Mateusz Maciejewski, Rama Harihara

Abstract: Deep learning-based denoising and upscaling techniques have emerged to enhance framerates for real-time rendering. A single neural network for joint denoising and upscaling offers the advantage of sharing parameters in the feature space, enabling efficient prediction of filter weights for both. However, it is still ongoing research to devise an efficient feature extraction neural network that uses different characteristics in inputs for the two combined problems. We propose a multi-branch, multi-scale feature extraction network for joint neural denoising and upscaling. The proposed multi-branch U-Net architecture is lightweight and effectively accounts for different characteristics in noisy color and noise-free aliased auxiliary buffers. Our technique produces superior quality denoising in a target resolution (4K), given noisy 1spp Monte Carlo renderings and auxiliary buffers in a low resolution (1080p), compared to the state-of-the-art methods.

External IDs:doi:10.1145/3728297