Temporal Reconstruction Network for Gradient-Domain Path Tracing

Aiming at the problem that most existing gradient domain path tracking methods focus on intra-frame reconstruction and lack the reuse of inter-frame information, resulting in difficulty in maintaining inter-frame consistency in rendering results, a gradient domain network reconstruction method combining historical frame information is proposed to address this issue. This method maps historical frame information to the current frame through motion vectors, and then achieves high-quality image reconstruction through the reconstruction network; In the reconstruction of the network, a loss function term based on temporal information was designed to achieve a balance between spatial reliability and temporal stability of the reconstruction results; Furthermore, a reweighting module was designed to achieve reasonable transmission of intra-frame reconstruction results between frames, improving both visual quality and temporal continuity. The experimental results show that compared with the mainstream gradient domain path tracking network reconstruction method GradNet, the proposed method achieves higher single frame reconstruction quality in multiple test scenes such as Bedroom, Classroom, and Bathroom. It can effectively eliminate noise while preserving more lighting details, and effectively eliminate the flicker problem of continuous frames. The RelMSE is reduced by 14.28% to ~32.56%, and the LPIPS error is reduced by up to 50.95%.