ZurichCV #12

Zan Gojcic (NVIDIA): From Reconstruction and Generation to Generative Reconstruction

Neural reconstruction has traditionally focused on faithfully recovering geometry and appearance from posed real-world images, while large generative models have excelled at synthesizing realistic content beyond the constraints of captured data. Recently, this distinction has begun to blur. Neural reconstruction methods are increasingly learned on large datasets and benefit from powerful generative priors, while the advances in conditioning on sparse observations and camera control have made video generative modeling strong candidates for novel view synthesis. This raises fundamental questions: Are reconstruction and generation on the brink of convergence? Do we still need explicit scene representations? And is reconstruction ultimately a strongly conditioned generative problem? Download the Slides

Vanessa Sklyarova (ETH): Modeling Realistic Hairstyles for Digital Human Avatars While 3D avatar creation has advanced rapidly, realistic hair remains a bottleneck. Current methods using meshes or implicit surfaces often lack the physical realism required for production pipelines. This talk presents a framework for physically plausible, strand-based hair modeling. I introduce methods for hair reconstruction from multi-view and monocular video, utilizing structured Gaussian splatting for high-frequency detail and synthetic priors for single-image reconstruction. Finally, I demonstrate conditional generative models that enable text-to-hair workflows, resulting in simulation-ready representations that bridge the gap between AI generation and production-grade rendering.