Stochastic progressive photon mapping (SPPM) is one of the important global illumination methods in computer graphics. It can simulate caustics and specular-diffuse-specular lighting effects efficiently. However, as a biased method, it always suffers from both bias and variance with limited iterations, and the bias and the variance bring multi-scale noises into SPPM renderings. Recent learning-based methods have shown great advantages on denoising unbiased Monte Carlo (MC) methods, but have not been leveraged for biased ones. In this paper, we present the first learning-based method specially designed for denoising-biased SPPM renderings. Firstly, to avoid conflicting denoising constraints, the radiance of final images is decomposed into two components: caustic and global. These two components are then denoised separately via a two-network framework. In each network, we employ a novel multi-residual block with two sizes of filters, which significantly improves the model’s capabilities, and makes it more suitable for multi-scale noises on both low-frequency and high-frequency areas. We also present a series of photon-related auxiliary features, to better handle noises while preserving illumination details, especially caustics. Compared with other state-of-the-art learning-based denoising methods that we apply to this problem, our method shows a higher denoising quality, which could efficiently denoise multi-scale noises while keeping sharp illuminations.