Retime GS turns choppy 2D clips into smooth 4D animations

Turning a casual smartphone clip into a dynamic 3D asset usually requires near-perfect inputs. The moment a video drops frames or stutters, most 4D reconstruction pipelines turn the result into a glitchy, ghosting mess.

Retime GS just demonstrated a practical fix. The system takes standard 2D video frames and generates smooth, full-body 4D scenes that viewers can explore from any angle.

Continuous motion over static snapshots

Existing dynamic reconstruction techniques like deform, gaussian flow, stgs, and 2D lift struggle when visual information is missing between frames. The resulting motion often twitches or breaks down with severe artifacts.

Retime GS solves this by introducing a continuous-time 3D representation. Instead of simply learning what a subject looks like at individual timestamps, the AI learns how motion evolves continuously over time. By pairing motion tracking and optical flow with an improved 4D Gaussian Splatting pipeline, it reliably interpolates missing movements and keeps dynamic bodies coherent.

Why it matters

We think Retime GS takes a pragmatic approach: rather than demanding pristine studio multi-cam setups, it adapts to messy real-world video clips.

For 3D creators, game animators, and VFX artists, this dramatically lowers the barrier to extracting usable motion assets from everyday monocular video. If continuous 4D representation becomes the norm, converting low-framerate footage into full 3D interactive assets will soon be routine.