Real-Time Neural Homogeneous Translucent Material Rendering Using Diffusion Blocks

This is the authors' abstract. We don't add key points for this paper.

Rendering realistic appearances of homogeneous translucent materials, such as milk and marble, poses challenges due to the complexity of subsurface scattering. In this paper, we present a neural method for real-time rendering of homogeneous translucent objects. Based on the observation that light propagation inside a highly scattered media is like a diffusion process (Stam 1995), we propose a neural data structure named diffusion block to mimic the behavior of the diffusion process. The diffusion block is built upon a recent network structure named DiffusionNet (Sharp et al. 2022) with a few modifications to adapt to our problem of translucent rendering. Our network is lightweight and efficient, leading to a real-time rendering method. Furthermore, our method supports dynamic material properties and diverse lighting conditions. Comparisons with state-of-the-art real-time translucent rendering methods demonstrate the superiority of our method in rendering quality.