See4D: Pose-Free 4D Generation via Auto-Regressive Video Inpainting
See4D synthesizes consistent 4D scenes from a single video without needing manual camera poses or trajectories. It replaces explicit motion prediction with rendering to fixed virtual cameras and uses a view-conditional inpainting model trained on depth-warped images. This pose-free approach separates scene modeling from camera control, enabling robust cross-view generation for applications like robotics and autonomous driving.