Neural Resampling with Optimized Candidate Allocation

We propose a novel framework that accelerates Monte Carlo rendering with the help of machine learning. Unlike previous works that learn parametric distributions that can be sampled directly, our method learns the 5-dimensional unnormalized incident radiance field and samples its product with the material response (BRDF) through Resampled Importance Sampling. This allows for more flexible network architectures that can be used to improve upon existing path guiding approaches and can also be reused for other tasks such as radiance caching. To reduce the cost of resampling, we derive optimized spatially-varying candidate counts to maximize the efficiency of the render process. We designed our method to accelerate CPU production renders by benefiting from otherwise idle GPU resources without need of intrusive changes to the renderer. We compare our approach against state-of-the-art path guiding methods, both neural and non-neural, and demonstrate significant variance reduction at equal render times on production scenes.