OuroWorld: Bringing Any 3D World Alive as Diverse, Endlessly Looping 3D Cinemagraphs
You-Zhe Xie, Ting-Wei Chou, Yu-Hsuan Li, Kaipeng Zhang, Zhixiang Wang, Yu-Lun Liu
OuroWorld converts static 3D scenes into seamlessly looping 3D cinemagraphs using periodic Fourier deformation and a drift-robust optimization.
How can we automatically transform static 3D Gaussian Splatting scenes into diverse, endlessly looping 3D cinemagraphs without manual masking?
Current 3D world models generate photorealistic scenes that remain frozen in time, while existing 3D cinemagraph methods are limited to fluid-like motion and require manual masks to define what moves. OuroWorld automates this by using a vision-language model to infer motion and a video diffusion model to generate multi-view dynamics, which are then fitted into a 4D Gaussian representation. The system uses a Fourier-series deformation field to guarantee seamless looping and a Grounded Drift Field to absorb inconsistencies across generated views. On 39 diverse scenes, OuroWorld outperforms all baselines in vividness and naturalness, winning up to 99% of user-study comparisons.
Paper Primer
The core mechanism hinges on two components: a Fourier-series deformation field that enforces periodicity by construction, and a Grounded Drift Field that isolates generative noise. The drift field acts like a shock absorber: it captures view-specific inconsistencies during training so the shared 4D scene remains stable and sharp, then is discarded entirely at inference.
OuroWorld achieves superior motion fidelity and loop coherence compared to Eulerian-based 3D cinemagraph methods.
User-study preference rates and ground-truth-free metrics (Vividness Degree and Motion-Aware Loop Fidelity). Preferred in 70.8%–99.0% of user-study comparisons across all baselines.
Why is a "Grounded Drift Field" necessary instead of just training on the generated videos directly?
Generated multi-view videos are inherently inconsistent; training directly on them forces the model to average conflicting observations, which blurs the scene and creates erratic motion. The drift field allows the model to attribute these inconsistencies to view-specific noise rather than the underlying 4D scene.
How does this approach differ from prior Eulerian-flow methods?
Eulerian methods are restricted to fluid-like motion and require manual masks to define moving regions. OuroWorld is mask-free and uses a Fourier-series basis to capture general deformation, object motion, and illumination changes that are not strictly fluid.
OuroWorld demonstrates that high-quality, mask-free 4D scene animation is possible by combining generative video priors with periodic mathematical constraints, effectively bypassing the need for ground-truth dynamic data.
Abstract and Overview
We outline the mask‑free pipeline, evaluation metrics, and superior performance on 3D cinemagraph generation.
Recent 3D world models generate photorealistic, explorable scenes that remain frozen in time, motivating the need for dynamic extensions.
OuroWorld is a mask‑free framework that transforms any static 3D Gaussian Splatting scene into a 3D cinemagraph, delivering vivid, diverse motion that loops seamlessly from any viewpoint.
A vision‑language model infers plausible dynamics and guides a video model to synthesize a reference video, which is then lifted and completed into multi‑view videos.
To learn from this imperfect supervision we propose Inconsistency‑Robust Periodic 4DGS: a Fourier‑series deformation field guarantees looping by construction, while a Grounded Drift Field anchored at the reference view absorbs cross‑view inconsistency.
Unlike prior Eulerian methods limited to fluid‑like motion, our approach captures general deformation, object motion, and illumination change.
We introduce a ground‑truth‑free evaluation covering vividness, naturalness, loop seam coherence, and scene quality, and demonstrate on 39 reconstructed and generated scenes that OuroWorld outperforms all baselines, winning 70.8%–99.0% of user‑study comparisons.
Introduction and Motivation
We expose why static 3DGS scenes need mask‑free, looping motion and outline our solution.
Static 3D world models such as HY‑World 2.0, Marble, and Lyra 2.0 render photorealistic scenes, yet everything is frozen: leaves do not sway, mist does not drift, lights do not flicker. Existing 3D cinemagraph approaches inherit Eulerian flow, can only express fluid‑like motion, and require manually provided masks. Video generation models produce diverse dynamics but their outputs are only approximately periodic and inconsistent across viewpoints, leading to seams and blurry motion when projected into 3D.
We need a mask‑free way to animate static 3DGS scenes with seamless looping motion that works from any viewpoint and supports arbitrary deformation, illumination change, and object motion.
3DGS represents a scene as a cloud of oriented 3D Gaussians whose splatted contributions are rasterized into a photorealistic image in real time.
Building on this motivation, we introduce OuroWorld, a mask‑free framework that turns any static 3DGS scene into a vivid, endlessly looping 3D cinemagraph.
**Figure 1** OuroWorld brings static 3D worlds alive. *Left*: Given a static 3DGS scene from any source, e.g., the 3D world models HY-World 2.0, Marble, and Lyra 2.0, OuroWorld converts it into a 3D cinemagraph: a dynamic scene with vivid, diverse motion that loops endlessly. *Right*: free-viewpoint renders along a moving camera, with **green boxes** highlighting regions with noticeable dynamics. Smoke billows in the volcano (top), the galaxy scatters stardust (middle), and the ships bob up and down on the water (bottom). Motion stays consistent across viewpoints, and the scene at $t=T$ returns to $t=0$, so the loop is seamless. See the supplementary videos for full results.
**Figure 2. Motivation.** *Left:* 3D world models generate frozen worlds; 3D cinemagraphs bring them alive. *Middle:* Existing 3D cinemagraph methods rely on Eulerian flow and are limited to fluid-like motion. *Right:* We turn any static 3D scene into a 3D cinemagraph, which requires (1) seamless looping and (2) vivid, diverse dynamics.
The key shift is turning a static 3DGS scene into a looping 3D cinemagraph without masks, enabling general motion and illumination changes.
The Periodic Deformation Field
We encode looping motion as a Fourier‑based offset field that repeats every period.
Static 3DGS scenes face three hurdles: diverse dynamics, loopable motion, and learning 4D from noisy multi‑view video supervision.
**Figure 3.** Overview of OuroWorld. (a) Looping video generation. A VLM selects a reference view rendered from the input 3DGS, infers plausible scene dynamics, and prompts a video generation model to synthesize an approximately looping reference video. (b) Multi-view video generation. A 3D foundation model (VGGT-$\Omega$) lifts the reference video into a dynamic point cloud, using static multi-view renders to improve depth. The point cloud is rendered from $N$ surrounding viewpoints into incomplete videos with disoccluded holes, which a video inpainting model (TrajectoryCrafter) completes into multi-view videos. (c) Inconsistency-Robust Periodic 4DGS. The 3D cinemagraph is represented as a canonical 3DGS deformed by a Fourier-parameterized Periodic Deformation Field, which loops by construction. A Grounded Drift Field absorbs cross-view inconsistency of the generated videos during training and is discarded at inference.
A smooth offset field that repeats every period $T$, so each Gaussian’s position, orientation, scale, and color are shifted in lockstep with a sinusoidal schedule.
How does this Periodic Deformation Field differ from earlier Fourier‑based dynamic representations?
Earlier works use Fourier bases merely as a flexible function approximator; they do not enforce $f(x,t+T)=f(x,t)$, so the resulting motion can drift and fail to loop. Our field explicitly builds the series with integer multiples of $2\pi/T$, guaranteeing exact periodicity by construction.
Compute $f(x,0)=0 + 1\cos(0) + 0\sin(0) + 0.5\cos(0) + 0.5\sin(0)=1.5$.
Compute $f(x,1)=0 + 1\cos(\pi/2) + 0\sin(\pi/2) + 0.5\cos(\pi) + 0.5\sin(\pi)=0 - 0.5 = -0.5$.
Compute $f(x,2)=0 + 1\cos(\pi) + 0\sin(\pi) + 0.5\cos(2\pi) + 0.5\sin(2\pi)= -1 + 0.5 = -0.5$.
Compute $f(x,3)=0 + 1\cos(3\pi/2) + 0\sin(3\pi/2) + 0.5\cos(3\pi) + 0.5\sin(3\pi)=0 - 0.5 = -0.5$.
Because $f(x,4)=f(x,0)=1.5$, the offset repeats after $4$ frames, guaranteeing a seamless loop.
The sinusoidal coefficients produce a periodic offset that returns exactly to its start after $T$ frames, illustrating why the field enforces loop‑consistency without any post‑hoc correction.
Scene-View Consistent Optimization
We isolate view‑specific motion with a Grounded Drift Field to keep all views dynamically consistent.
Multi‑view videos generated from a single 3DGS often disagree: each view sees a slightly different motion, so naïvely fitting them forces the deformation field to average contradictory cues, producing blur and jittery motion.
We give every view its own tiny, time‑varying offset $\Delta$ that absorbs view‑specific errors, but we force the canonical (reference) view to rely solely on the shared scene $P$, so the scene itself must learn the true dynamics.
Render reference view 0 at $t=0$: $G + P = $ Gaussian at $(0,0,0)$, no drift.
Render reference view 0 at $t=1$: $G + P = $ Gaussian at $(+1,0,0)$, no drift.
Render alternate view 1 at $t=0$: $G + P + \Delta_1 = $ Gaussian at $(0,0,0)$ with original color.
Render alternate view 1 at $t=1$: $G + P + \Delta_1 = $ Gaussian at $(+1,0,0)$ with color $+0.2$ shift.
During training the loss attributes the color mismatch at $t=1$ in view 1 to $\Delta_1$, leaving $P$ unchanged.
At inference we drop $\Delta_1$, so both views share the same motion (pure translation) and the color remains consistent.
Grounding forces the shared deformation $P$ to capture the true motion, while $\Delta$ only mops up view‑specific artefacts; discarding $\Delta$ at test time yields a clean, consistent animation.
How does this Grounded Drift Field differ from the per‑view drift used in Video‑to‑World?
Video‑to‑World adds a drift that is learned after aligning geometry with ICP, but the drift is applied to every view—including the reference—so the shared scene can become static. Our Grounded Drift Field applies drift only to non‑reference views and explicitly forces the reference view to rely on $P$ alone, preventing the trivial solution where drift explains all dynamics.
Even with a grounded drift, the generated frames may still diverge from the original 3DGS. To tighten this gap we supervise the most reliable images—the reference video $V_{\text{ref}}$ and the $t\!=\!0$ multi‑view renders of the input 3DGS—using an $\ell_1$ loss $\mathcal{L}_1$.
Compute $\mathcal{L}_1$ between the reference video $V_{\text{ref}}$ and the rendered $t\!=\!0$ views of the input 3DGS.
For each remaining generated frame, add Gaussian noise to the rendering.
Run a Stable Diffusion v1.5 image diffusion model, guided by the noisy frame, to obtain a refined image.
Measure the perceptual similarity between the original rendering and the refined image with LPIPS, yielding $\mathcal{L}_p$.
Combine the two losses: $\mathcal{L}= \lambda_1 \mathcal{L}_1 + \lambda_p \mathcal{L}_p$.
**Figure 4** Learning deformation from inconsistent multi-view videos. Rows: canonical geometry, deformed scene at inference, per-view training renders, and generated training views. (a) w/o Drift Field: conflicting views are averaged, causing blur and erratic motion. (b) w/ Drift Field: drift on every view absorbs too much of the dynamics and is discarded at inference, reducing motion. (c) w/ Grounded Drift Field (ours): with no drift at the reference view, the deformation field must reproduce the reference dynamics, yielding a sharp scene with consistent motion.
Experimental Evaluation
We compare baseline methods on vividness and loop consistency using our ground‑truth‑free metrics.
We evaluate four existing 3D cinemagraph approaches alongside our method, applying identical input preprocessing and rendering pipelines to ensure a fair comparison.
These baselines each generate dynamic 3D scenes from a static 3DGS, but differ in how they model motion and whether they enforce loop consistency.
**Figure 5** Qualitative comparison under the orbit camera, one scene per source. The leftmost column shows the input 3DGS at $t_1$; the others show each method at later time steps $t_2$ and $t_3$. 3D Cinemagraphy and 3D-MOM produce disocclusion holes and color artifacts, while Gaussians-to-Life and LoopGaussian remain nearly static. OuroWorld produces vivid, view-consistent dynamics such as changing illumination, rippling water, drifting fog, and curtains swaying in the wind. See the supplementary videos for full loops.
OuroWorld outperforms existing methods in both vividness and loop consistency.
Quantitative Results
Quantitative comparison of our method against baselines across multiple metrics.
Our method achieves the highest Vividness Degree among all baselines.
Table 2 shows Ours at 0.0423, exceeding the next best 0.0374 (Gaussians‑to‑Life).
**Figure 6. User study.** In a two-alternative forced-choice study, participants compare OuroWorld with each baseline under four criteria. Bars show the percentage of votes preferring ours (left) versus the baseline (right). Ours is preferred in 70.8%–99.0% of comparisons across all baselines and criteria, with the largest margins in Vividness. 30 participants took part; see Appendix A.6 for details.
Metrics and User Study
We evaluate our method with KVD, loop fidelity, and scene‑quality metrics, showing top performance.
KVD quantifies how far a generated video distribution deviates from natural video distributions using a kernel‑based similarity measure.
These metrics assess whether a generated video returns to its initial state smoothly after each loop, ensuring temporal continuity.
Table 2 shows that our method attains the highest scores on Vividness, Naturalness (KVD), MALF, and Aesthetic Quality for both static and orbit camera settings.
Qualitative inspection (Figure 5) reveals that competing methods either produce static outputs or exhibit disocclusion artifacts, while our approach delivers vivid dynamics that remain consistent across novel viewpoints.
The user study with 30 participants shows a clear preference for our results over all baselines on Vividness, Naturalness, Loop Seam Coherence, and Visual Quality, with the largest margin on Vividness.
Ablation experiments (Table 3) indicate that removing the Grounded Drift Field, the scene‑view consistent optimization, or the periodic deformation period degrades Vividness, consistency, and sharpness, confirming the importance of each component.
**Figure A.2 Motion masks.** (a) The 2D motion mask obtained by segmenting the moving entities in the reference image with SAM 3. (b) The corresponding 3D motion mask.
We conclude that OuroWorld reliably turns static 3DGS scenes into looping 3D cinemagraphs with diverse motion, while noting that the dynamics are bounded by the video‑generation prior and that single‑view reference limits view‑wide consistency.
The appendix provides full implementation details, hyperparameters (Table 4), baseline inputs (Table 5), vividness breakdowns (Table 6), ablation comparisons (Table 7), and dataset‑collection procedures.
Ablations and Limitations
Ablation analysis quantifies each component’s impact and details the user‑study evaluation.
We isolate the contribution of each design element by removing it from the full pipeline and measuring the resulting drop in key quality metrics.
Removing the drift field blurs the scene (lower VoL) and reduces subject consistency.
Table A.4 shows subject consistency falls from 0.0588 to 0.0374, a ≈ 36 % drop, while VoL drops by ≈ 22 % under the orbit‑camera setup.
Removing grounding lets the drift field absorb dynamics, cutting overall motion and vividness.
Vividness drops from 0.0423 (full model) to 0.0247 when grounding is omitted, a ≈ 41 % reduction.
Removing scene‑view consistent optimization breaks view‑consistent motion, further lowering subject consistency and VoL.
Under the orbit camera, subject consistency falls to 0.0259 and VoL declines by an additional ≈ 22 % compared with the full model.
All three components—drift, grounding, and scene‑view consistent optimization—are essential; the full model outperforms each ablated variant across both static and orbit viewpoints.
Naturalness follows the same pattern: the full model achieves the highest KVD‑based naturalness scores, while each ablation reduces the metric by 5–12 %.
We also evaluate temporal coherence with two complementary metrics. Motion‑Aware Loop Fidelity (MALF) measures how well a video returns to the same state after one period while penalizing static loops. Seam SSIM directly compares the last frame of one cycle with the first frame of the next, which can be artificially high for near‑static videos.
The user study uses a two‑alternative forced‑choice protocol: participants view side‑by‑side videos (our method vs. a baseline) and answer four criteria—Vividness, Naturalness, Loop Seam Coherence, and Visual Quality.
Across 39 scenes and four baselines, each video pair was judged by five participants, yielding 780 trials and 780 votes per criterion (195 votes per baseline). Preference rates are reported in Fig. 6.
**Figure A.3** User study interface. Two videos of the same scene are shown side by side in random order (top), and participants choose one for each of the four criteria (bottom).