Teaching 3D scenes to handle tricky camera angles

Urban scene reconstruction from camera footage works great along recorded paths but falls apart when the viewpoint shifts dramatically—a critical problem for autonomous vehicle simulation. GenRe solves this by using diffusion models to distill generative priors into existing 3D Gaussian representations, improving quality and robustness in under 5 minutes per scene. Unlike prior methods, it generalizes reliably to unseen extreme angles (like lane changes) across diverse urban environments, enabling more robust closed-loop simulation for self-driving development.