Real‐time Monte Carlo Denoising with Weight Sharing Kernel Prediction Network

Real-time Monte Carlo denoising aims at removing severe noise under low samples per pixel (spp) in a strict time budget. Recently, kernel-prediction methods use neural network to predict each pixel's filtering kernel and have shown great potential remove noise. However, the heavy computation overhead blocks these from real-time applications. This paper expands method proposes novel approach denoise very spp (e.g., 1-spp) path traced images frame rates. Instead of using directly map, i.e., complete weights per-pixel kernel, we an encoding followed by high-efficiency decoder with unfolding operations for high-quality reconstruction kernels. The map yields compact single-channel representation which can significantly reduce network's throughput. In addition, adopt scalable fusion module improve quality. proposed preserves prediction methods' quality while roughly halving its 1-spp noisy inputs. compared recent bilateral grid-based denoiser, our benefits high parallelism kernel-based produces better results equal time.