Reconstructing moving scenes from video without knowing camera angles

NoPo4D solves a gap in 3D scene reconstruction: prior methods handle dynamics with known poses, multi-view with static scenes, or require expensive per-scene optimization. This feed-forward approach splits Gaussian motion into 2D image shifts and depth changes, enabling direct supervision from optical flow rather than 3D motion ground truth, plus bidirectional cross-view feature blending. On four benchmarks, it outperforms feed-forward baselines and matches optimization methods while running much faster.