Abstract: This talk explores three distinct technical paradigms for generative vision models in 3D reconstruction and synthesis. First, we present a novel 3D rectified flow matching model trained from scratch for robotic assembly, demonstrating how flow-based trajectories can be optimized for precise geometric reasoning. Second, we discuss the architectural adaptation of video diffusion models to enhance 3D Gaussian Splatting (3DGS); by integrating specialized encoding modules into a foundation model that leverage 3DGS priors, we bridge the gap between 2D temporality and 3D spatial consistency. Finally, we introduce a test-time optimization technique for 3D style transfer that utilizes pretrained large 3D generative models to align disparate geometries. Together, these works illustrate a versatile toolkit for modern 3D vision—from designing specialized generative flows to the sophisticated manipulation of large-scale latent priors.

Host: Prof. Francis Engelmann