Should generative models focus harder on uncertain LiDAR regions?

Generating realistic 4D scenes from LiDAR scans requires different modeling effort for different regions—distant surfaces and occluded boundaries are inherently harder to reconstruct than well-observed areas. U4D allocates generation capacity strategically using per-point uncertainty maps derived from segmentation entropy: it first generates high-uncertainty regions precisely via diffusion, then conditionally fills the rest. A spatio-temporal mixer balances geometric detail with frame-to-frame consistency, improving both scene fidelity and downstream task performance.