Open-Source CPU-Runnable Reimplementation of Neural Render Proxies (EGSR 2026) Released

A sample reimplementation of "Neural Render Proxies for Interactive and Differentiable Lighting" (Sancho et al., Computer Graphics Forum / EGSR 2026) has been released as an open-source, educational project. The core idea from the paper is to decouple path tracing into a light-agnostic path pass (SAMPLEPATHS) and a cheap emission pass (GATHERLIGHT), with a compact per-light-type MLP compressing the path cache into a differentiable proxy — enabling interactive relighting in the forward pass and gradient-based inverse lighting via backpropagation.

The reimplementation ships two backends: a dependency-light NumPy reference with hand-rolled autodiff, and a PyTorch backend matching the paper's architecture including 2D multiresolution hash encoding and the paper's inverse-optimization formulation. Everything runs end-to-end on a laptop CPU in minutes with no GPU required, though GPU acceleration via MPS is supported and achieves 5–7× speedups. All ten roadmap items are reported complete, with a paper-scale MLP reaching 35.19 dB held-out PSNR on a Mitsuba Cornell box scene.