Neural Render Proxies for Interactive and Differentiable Lighting
Computer Science · arXiv:2611.17185
Pith says
Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded
Holds up, with conditions. Clean idea, sound math, but the headline claim of arbitrary light placement is not tested: validation lights come from the training distribution, and the paper itself shows the failure mode in under-sampled regions.
4 major objections, the first at §5.1, Figure 6, and §4.4 · 10 minor