Adventures in Neural Rendering

A graphics programmer explores using small multilayer perceptrons (MLPs) to encode rendering data like radiance, irradiance, depth, ambient occlusion, and BRDF. The experiments compare MLP-based encoding against traditional methods like Spherical Harmonics, finding that MLPs can encode radiance efficiently with fewer parameters than L2 SH, but struggle with irradiance and specular BRDF representation. The Rusinkiewicz parameterization significantly improves BRDF encoding. While MLPs show promise for signal encoding in rendering, training time and inference costs remain significant challenges for real-time applications.