Accelerated Path Tracing With GAN and Matrix Completion

Denoising-based techniques have recently been shown to be effective for accelerating path tracing rendering methods. However, there remains a problem which is input images need the minimum necessary samples number in order to ensure the quality of the output. In this paper, we propose a new accelerated path tracing approach with generative adversarial networks(GAN) and matrix completion. Unlike the methods based on denoising with neural network, we randomly render part of pixels of input image, which are much less than other methods. Next, we utilize the trained GAN to pre-complete the initializing missing pixels. Because of the accuracy and fast-convergence of GAN, our pre-completion results are more accurate than other methods. Then, according to the results of pre-completion, we present the pre-completed images as a low-rank matrix and make use of the matrix completion to recovers missing values accurately even in high details. To improve the efficiency of solving matrix completion, we modified the original weighted nuclear norm minimization with a parameter adjustment(PAWNNM) strategy. The result shows better visual quality, texture details and convergence efficiency than the state-of-the-art acceleration methods, especially the methods based on denoising with neural network.