OuroWorld: Bringing Any 3D World Alive as Diverse, Endlessly Looping 3D Cinemagraphs

OuroWorld. Bringing Any 3D World Alive as Diverse, Endlessly Looping 3D Cinemagraphs. By You-Zhe Xie and five others, from National Yang Ming Chiao Tung University and Alaya Lab. Published on the eighth of October.

Recent 3D world models can generate scenes that are photorealistic and can be explored, yet those scenes are frozen in time. Nothing in them moves. Water does not flow and light does not change. Adding motion is harder than it sounds, because the motion has to look right from every angle a viewer might choose, and for a cinemagraph it also has to come back to where it started without a visible jump.

OuroWorld turns any static 3D Gaussian Splatting scene into a 3D cinemagraph: a dynamic scene with vivid and diverse motion that loops seamlessly, seen from any viewpoint. It is mask-free, so nobody has to mark which parts should move.

A vision-language model first infers what dynamics would be plausible for the scene, and guides a video model to synthesize a reference video. That video is lifted and completed into multi-view videos.

Those videos are imperfect supervision, and learning from them is the job of Inconsistency-Robust Periodic 4DGS. It has two parts. A deformation field built from a Fourier series guarantees the loop by construction. A Grounded Drift Field, anchored at the reference view, absorbs the inconsistencies between views. Earlier Eulerian methods were limited to motion that behaves like fluid. This approach captures general deformation, object motion, and changes in illumination.

Because there is no ground truth for such scenes, the authors introduce an evaluation that needs none. It covers vividness, naturalness, coherence at the loop seam, and scene quality. On thirty-nine reconstructed and generated scenes, OuroWorld outperforms every baseline, and wins between seventy point eight and ninety-nine percent of user-study comparisons.

Looping is not something the method hopes to learn. It is built into the motion model itself, while a separate field soaks up the errors that the generated supervision brings with it.