Abstract: In this talk, I will talk about the promise---and challenges---of moving neural primitives closer to the core of how we generate images, from the perspective of select recent research from NVIDIA's real-time graphics research group.

Advances in machine learning techniques have fundamentally transformed how we solve many challenging computational problems, but we have yet to see their full impact in rendering. Although renderers have made successful use of deep learning for years in the form of denoisers and upsamplers, there is still a distinct boundary between the (largely traditional) renderer and the (now largely neural) post-process denoiser. Bridging that gap and moving neural models into the renderer itself represents a major shift in how we generate images, and simultaneously brings new tools to tackle long-unsolved challenges, and brand-new challenges that need solving of their own before these methods are viable---extending all the way down to the programming languages we use to build our renderers.

Bio: Benedikt is a Senior Research Scientist at NIVIDA, where he works on problems at the intersection of graphics and machine learning for the post-Moore's law era. In the past, Benedikt worked on topics including light transport, appearance modeling and denoising, primarily centered on offline rendering. Since publishing ReSTIR, Benedikt is increasingly focused on topics in real-time, with a particular focus on appearance modeling and scene representation using neural models.Benedikt holds a PhD degree from Dartmouth College (2022), and a Master’s (2015) and Bachelor’s degree (2013) from ETH Zurich. Away from the computer, he is likely found in his workshop, or in the wild collecting samples for his microscopes.