How to reconstruct moving scenes from video without per-frame geometry?

Reconstructing dynamic scenes from single-camera video typically requires predicting 3D geometry separately for each frame, leading to duplicate structures and poor motion understanding. C4G instead learns a compact set of reusable Gaussian tokens conditioned on timestamp, letting each token aggregate temporal context and shift position over time for globally coherent motion. A diffusion-based rendering enhancer captures fine details, and the same framework extends to 4D feature fields for point tracking—all without camera pose requirements or per-scene optimization.