ZurichCV #14

Haithem Turki (Senior Research Scientist, NVIDIA): Photoreal Worlds at Scale – From Capture to Generation

Neural reconstruction and generative video modeling have evolved in parallel: the former focused on faithfully recovering photoreal 3D scenes from real-world sensors, the latter on synthesizing plausible content at internet scale. Recently, the boundary between them has begun to blur. Per-scene reconstruction is bounded by what was observed, motivating generative priors to fill in unseen regions, while large video models — strong on plausibility but lacking explicit geometry — are increasingly conditioned on 3D scaffolding to enforce spatial consistency. This talk walks through three recent works at this intersection, and asks: How should explicit 3D representations handle real sensor complexity? Where do generative priors fit without overwriting what was captured? And is throwing more parameters at 3D foundation models really the answer?

Omar Sanseviero (Google DeepMind): Gemma, DeepMinds's family of open models Running open-weights models entirely locally tackles data privacy, latency, and cloud cost issues, but it requires extreme efficiency. In this session, we explore the design philosophy behind Google's Gemma 4, focusing on how we pushed the boundaries of intelligence per parameter for edge hardware. Beyond just running the model, we'll look under the hood at the model quantization, hardware optimization, and architectural decisions required to build high-performance, on-device AI.