Abstract:When using Monte Carlo path tracing method to render 3D scenes, artifacts may occur due to insufficient sampling. Di-rectly increasing the number of samples can increase the time cost of the rendering algorithm. An effective strategy is to raise sampling levels iteratively. However, iterative operations introduce additional computational overhead. To solve this problem, we proposed a rendering acceleration method that does not require iterative computations. This method com-bined extraction of the Just Noticeable Difference (JND) information and used a neural network to predict the sampling matrix of the scene, which was adjusted based on the lighting information in the pre-rendered image. To start with, we extracted the JND information from pre-rendered images and estimated the rapid convergence regions, such as environ-ment mapping regions and light source regions. We then employed the Conv-LSTM to estimate the JND features for high-quality rendered images. We then employed the Conv-LSTM to estimate the JND features for high-quality rendered images. We designed a multi-feature fusion network to predict the required number of samples for each pixel. The en-coder took the pre-rendered images as input, which were then fused with the JND features for the decoder to generate the corresponding sampling matrix. In addition, based on the observation about the slow convergence at low lighting areas, we adjusted the sampling matrix according to the lighting clustering results derived from the pre-rendered images. The experimental results indicated that our method has better performance compared with the current methods.