Currently, the main idea of iterative rendering methods is to allocate a fixed number of samples to pixels that have not been fully rendered by calculating the completion rate. It is obvious that this strategy ignores the changes in pixel values during the previous rendering process, which may result in additional iterative operations. To solve this problem, we propose a sampling allocation method based on just noticeable difference (JND) and sample gradient information. Firstly, we obtain the regions in the scene with faster convergence speed by calculating the differences between four sets of pre rendered images, such as environment map regions, light source regions, etc. Afterwards, we use long short term memory (LSTM) to predict the JND information of high-quality rendering results, recorded as JNDf. Moreover, during the iterative rendering process, we use JNDf and sample gradients to calculate the number of additional samples required for the next iteration. Finally, we use the fixed JND threshold and grayscale change rate as the termination conditions for rendering. After experimental verification, our method’s rendering results have significant advantages in peak signal-to-noise ratio (PSNR) and structural similarity index measurement (SSIM).