OuroWorld
A mask-free framework that converts an arbitrary static 3D scene, such as a 3DGS reconstruction, into a dynamic 3D cinemagraph that loops seamlessly when viewed from any viewpoint. It uses a VLM to infer dynamics and a video model to synthesize a reference video, then promotes the result to a multi-view video sequence using an inconsistency-robust periodic 4D Gaussian splatting model.
Timeline (last 90 days)
OuroWorld is a mask-free framework capable of turning any static 3D scene into a dynamic image (3D cinemagraph) that loops seamlessly when viewed from any viewpoint.
OuroWorld's core method is to use a vision-language model (VLM) to infer scene dynamics, leverage a video model to synthesize a reference video, and then lift and complete it into a multi-view video sequence.
The research team tested OuroWorld on 39 reconstruction and generation scenarios, achieving leading results in all baseline comparisons and winning 70.8% to 99.0% of the win rate in user study comparisons.
OuroWorld proposes Inconsistency-Robust Periodic 4DGS, which contains two key components, the Fourier series deformation field and the Grounded Drift Field, to learn dynamics from imperfect supervision signals.
OuroWorld introduces an evaluation system that requires no ground truth, covering four dimensions: vividness, naturalness, loop seam coherence and scene quality.
OuroWorld adopts a Lagrangian-style deformation field framework rather than previous Eulerian-viewpoint-based methods, and combines VLM high-level semantic understanding to guide motion synthesis, covering a broader range of dynamic types (including rigid body motion, non-fluid deformation and illumination changes).
Most dynamic 3D methods attempt to align the first and last frames in a post-training processing stage to achieve looping, whereas OuroWorld embeds periodicity directly into the mathematical form of the deformation field, guaranteeing looping holds from the source.
The OuroWorld project homepage is https://ouroworld.userwei.com, containing looping dynamic demonstrations of multiple scenes, baseline comparisons and user study data.
All Facts (8)
OuroWorld is a mask-free framework capable of turning any static 3D scene into a dynamic image (3D cinemagraph) that loops seamlessly when viewed from any viewpoint.
50%UnverifiedOuroWorld's core method is to use a vision-language model (VLM) to infer scene dynamics, leverage a video model to synthesize a reference video, and then lift and complete it into a multi-view video sequence.
50%UnverifiedThe research team tested OuroWorld on 39 reconstruction and generation scenarios, achieving leading results in all baseline comparisons and winning 70.8% to 99.0% of the win rate in user study comparisons.
50%UnverifiedOuroWorld proposes Inconsistency-Robust Periodic 4DGS, which contains two key components, the Fourier series deformation field and the Grounded Drift Field, to learn dynamics from imperfect supervision signals.
50%UnverifiedOuroWorld introduces an evaluation system that requires no ground truth, covering four dimensions: vividness, naturalness, loop seam coherence and scene quality.
50%UnverifiedOuroWorld adopts a Lagrangian-style deformation field framework rather than previous Eulerian-viewpoint-based methods, and combines VLM high-level semantic understanding to guide motion synthesis, covering a broader range of dynamic types (including rigid body motion, non-fluid deformation and illumination changes).
50%UnverifiedMost dynamic 3D methods attempt to align the first and last frames in a post-training processing stage to achieve looping, whereas OuroWorld embeds periodicity directly into the mathematical form of the deformation field, guaranteeing looping holds from the source.
50%UnverifiedThe OuroWorld project homepage is https://ouroworld.userwei.com, containing looping dynamic demonstrations of multiple scenes, baseline comparisons and user study data.
50%