Taming optimization variance in compact neural shading networks
Auto-discovered
Visit resource ↗Description
Taming optimization variance in compact neural shading networks is a SIGGRAPH 2026 conference paper by Benedikt Bitterli, Petrik Clarberg, Chris Cummings, Aaron Lefohn, Steve Marschner, Jan Novák, Fabrice Rousselle, Andrea Weidlich, and Tizian Zeltner from NVIDIA Real-Time Graphics Research. The work addresses unstable training in small MLPs used for real-time neural shading, where different random seeds often land in unequal local minima.
The authors propose multi-instance training that maintains an ensemble of networks, prunes underperforming instances, and dynamically resizes batches to keep wall-clock cost near single-instance training while exploring weight space more thoroughly. Experiments focus on learning high-dynamic-range reflectance functions, a setting where tiny networks struggle with contrast and specular detail.
The paper also revisits loss functions, activations, and input parameterizations that improve robustness for compact architectures. Graphics programmers implementing neural materials in engines will recognize the trade space between network size, frame budget, and temporal stability.
NVIDIA hosts abstracts, PDF links, supplementary material, and code pointers on its Real-Time Graphics Research publication page. Content is free to read from that landing page for practitioners tracking neural shading deployments.
Bitterli and colleagues' neural shading publications are widely cited in real-time rendering research circles; this SIGGRAPH 2026 paper extends their line of work on production-viable learned materials.
Link last verified accessible on 2026-09-27.