Adventures in Neural Rendering
Summary
The article surveys early experiments using small MLPs to encode rendering signals (radiance, irradiance, depth) and compares them with Spherical Harmonics, highlighting trade-offs in storage, directional fidelity, and training/inference cost. It demonstrates that while MLPs can match or exceed SH storage for some cases, practical real-time use remains limited without shader-accelerated inference, and discusses parameterisations like Rusinkiewicz to improve BRDF representation.