acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Recent 3D world models generate photorealistic, explorable scenes that remain frozen in time. OuroWorld is a mask-free framework that turns any static 3D Gaussian Splatting scene into a 3D cinemagraph: a dynamic scene with vivid, diverse motion looping seamlessly from any viewpoint. A vision-language model infers plausible dynamics and guides a video model to synthesize a reference video, which we lift and complete 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, we capture general deformation, ob- ject motion, and illumination change. We introduce a ground-truth-free evaluation covering vividness, naturalness, loop seam coherence, and scene quality. On 39 reconstructed and generated scenes, OuroWorld outperforms all baselines and wins 70.8%–99.0% of user-study comparisons.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D cinemagraph, 4D Gaussian Splatting, dynamic scene representation, video generation prior, seamless looping motion, multiview inconsistency

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_GxtgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This manuscript presents a method for converting a static 3DGS scene into a 3D cinemagraph with diverse dynamics, without relying on masks or physical simulation. Given a reference image rendered from the 3DGS, the method first uses a proprietary LLM and video generation model to produce an approximately looping reference video. This video is then lifted and projected to multi-views, and inpainted as multi-view videos. The original 3DGS renderings, the reference video, and the inpainted multi-view videos are used to train the proposed Inconsistency-Robust Periodic 4DGS, where the Periodic Deformation Field is used to capture periodic motion, the Grounded Drift Field is designed to absorb inconsistencies among the generated videos during training. When optimizing the reference view, the Drift Field is disabled to encourage the Periodic Deformation Field to explain the desired periodic motion. This manuscript also introduces a perceptually-aligned ground-truth-free evaluation framework for 3D cinemagraph, under which the proposed approach shows a clear advantage over the alternatives.

Critical strengths
  1. This work explores an interesting problem, and the generated videos are overall satisfactory.
Critical weaknesses
  1. Grounding the drift field sounds reasonable in principle. But in many scenes, anchoring a single view is insufficient to determine the its 3D and motions, so it is unclear whether this mechanism can work as expected. So in addition to the quantitative evaluation, it is suggested to include some qualitative ablation about the grounding mechanism, preferably in video. Moreover, if the goal is to enlarge the motion amplitude modeled by Periodic Deformation Field, an appropriate regularization term might already be sufficient. It would be useful to include such baselines.
  2. The merit and necessity of the mask-free need further justification. Based on the ablation videos in the supplementary material, the improvement brought by the drift field seems to be happened mainly in static regions. Similar improvements could potentially be achieved at lower cost by just using a mask during optimization.
Questions for authors
  1. Why not directly use Fourier series to model the periodic attributes instead of the features?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Gxtg

We thank the reviewer for recognizing (1) an interesting problem and (2) satisfying generated videos. We address both concerns and the question below.

W1 — Whether grounding to a single view is enough.

What grounding does. A per-view Drift Field absorbs the disagreement between generated views during training and is discarded at inference. Grounding removes the drift at the reference view, so the drift cannot take over the scene motion.

Qualitative ablation. We added w/o Grounding vs. Ours videos at the bottom of the supplementary index.html. Without grounding, the drift absorbs scene structure and details (City Explosion) or the intended motion, such as the lake ripples in Mountain Lake, instead of only cross-view disagreement.

Limitation. We agree grounding anchors the drift at one view only. If some training views differ from it by a very large angle, it may not keep the motion out of their drift. We will state this in the limitations.

Drift-magnitude regularizer instead of grounding. As suggested, we replace grounding with a penalty λ‖Δ‖² on the drift (λ = 10³), applied to all views including the reference; all else unchanged. D/M measures drift size relative to the deformation, Vividness how much the scene moves, and VoL sharpness. All 39 scenes, Table 2 protocol:

CameraMethodD/MVividness ↑VoL ↑
Staticw/o Drift–0.0424 (+0.2%)883.55 (−1.0%)
Staticw/o Grounding1.8630.0392 (−7.3%)882.44 (−1.1%)
StaticDrift + magnitude reg.0.0250.0408 (−3.6%)882.92 (−1.1%)
StaticOurs2.0780.0423892.43
Orbitw/o Drift–0.5113 (−0.1%)704.34 (−0.8%)
Orbitw/o Grounding1.8630.5080 (−0.7%)681.16 (−4.0%)
OrbitDrift + magnitude reg.0.0250.5086 (−0.6%)671.68 (−5.4%)
OrbitOurs2.0780.5117709.75

*D/M: drift translation / deformation translation, averaged over views. Percentages relative to Ours.

  • Unconstrained drift absorbs motion; the regularizer recovers only half. Without grounding, static Vividness drops 7.3%; the penalty lifts it from 0.0392 to 0.0408, still 3.6% below Ours.
  • It does so by switching the drift off. D/M falls from 2.078 to 0.025, so the drift no longer absorbs cross-view disagreement, and orbit renderings blur (VoL −5.4%). The penalty acts on all drift equally, so it cannot tell drift that absorbs disagreement from drift that steals motion.
  • Grounding keeps both. The drift stays large (D/M 2.078), yet Vividness stays high and VoL is highest under both cameras: grounding blocks only the harmful case. Under orbit, camera motion dominates Vividness, but the ungrounded variants lose 4.0–5.4% VoL. We will add this baseline to the paper.

W2 — Why mask-free, instead of a motion mask during optimization.

  • Masks cannot isolate many target dynamics. Lighting changes, shadows, smoke and fog have no clean object boundary, yet they are key effects OuroWorld aims for.
  • Accurate 3D masks are hard even for simple object motion. We explored this route. Back-projecting SAM 3 masks into 3D needs a depth range, and none works across scenes. Too narrow, and only visible-surface Gaussians are selected; interior and occluded ones stay frozen, so the surface "peels" off the moving object (video at the bottom of the supplementary index.html; Fig. 12 in the revision). Too wide, and static regions are included, bringing the artifacts back.
  • Masks are not cheaper in practice. In our baseline pipeline (App. D.2), every mask needed human verification; OuroWorld needs no manual input.

Q1 — Why put the Fourier series on triplane features rather than per-Gaussian attributes?

Following 4DGS, which models motion with a HexPlane, we place the Fourier coefficients on a triplane of the same form. Compared with per-Gaussian coefficients, this has three advantages:

  • Spatial coherence. Neighboring Gaussians query shared plane features, so their periodic motion and color vary smoothly in space; per-Gaussian coefficients have no such coupling.
  • Scalability. The triplane has a fixed size (39.0M parameters) regardless of Gaussian count, while per-Gaussian coefficients grow linearly with it, needing 3.3–76× more dynamics parameters:
Dynamics stored onDynamics paramsTotal paramsPeak train VRAMPeak render VRAM
Triplane (ours)39.0M56M–400M~9 GiB~3 GiB
Per-Gaussian129M–2,960M146M–3,321M~9–54 GiB~3–15 GiB
  • Expressiveness. The Fourier series acts on features, and one non-linear MLP decodes them into position, rotation, scale and color. K thus limits only how fast the input feature changes, not the final motion: the MLP can turn a smooth input into a sharp change, such as a light switching on and off.

We thank the reviewer again, and are glad to answer further within the remaining discussion time.

Sincerely, The Authors

I have reviewed the rebuttal and supplementary materials and been satisfied with the newly provided results and discussions. For W2, while I agree that the authors' argument sounds overall reasonable, it cannot directly address my original concern stemminng from an observation of the video ablation in supplementary material, i.e., the improvement in the presented case happened mainly in static regions. In this case, we don’t really need an “accurate 3D mask”, a coarse and conservative mask would be already enough to eliminate artifacts in regions far from the dynamic areas. Overall, I decide to maintain my original score.

AuthorsOpenReview ↗

Follow-up on W2: Drift Field also matters in dynamic regions

We thank the reviewer for the constructive feedback. We observe that the effect of removing the Grounded Drift Field is not limited to static regions: dynamic regions also degrade, as the Marble/world scene of the supplementary ablation shows:

  1. Within the motion. The rotating galaxy is noticeably blurrier w/o Drift. The inconsistency lies in the moving content itself, so masking cannot help.
  2. Along the motion path. As the light sweeps across the stones, w/o Drift averages them out until they nearly vanish. A coarse mask over the lighting change will covers these stones, so they cannot be frozen as static, and the mask cannot prevent them from vanishing.

To quantify this, we measure sharpness (VoL) only within the dynamic region of each video, defined as the union of the pixels flagged by the motion and illumination terms of our Vividness metric (App. E).

Dynamic-region VoL ↑
w/o Drift938.10 (−1.43%)
Ours951.72

In some scenes the drop is much larger: World from Marble (−26%), City Entrance from Marble (−25%), Great Wall from Marble (−17%) and Snowy Riverbank from Lyra (-9%).

A mask can only suppress artifacts away from the motion, and cannot isolate boundaryless effects such as smoke, fog, and lighting changes. A reasonable coarse mask is also hard to obtain: the motion region is unknown before generation, and the dilation margin needs per-scene tuning. The Drift Field handles artifacts within, along, and away from the motion, without any mask.

We sincerely thank the reviewer for pushing the discussion on W2 further, which helped us clarify the strengths and limitations of our approach relative to mask-based methods.

Reviewer_P3NJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OuroWorld animates static 3D Gaussian Splatting scenes into diverse, seamlessly looping 3D cinemagraphs. It combines video-generated multi-view supervision with a Fourier-based deformation field that guarantees periodicity and a reference-anchored drift field that handles cross-view inconsistencies. Experiments on 39 reconstructed and generated scenes report improvements over four baselines in overall vividness, loop fidelity, and perceptual quality, supported by a user study.

Critical strengths
  • Practical framework. OuroWorld integrates video priors with 3D scene animation to produce diverse motion and appearance changes without user-provided motion masks.
  • Guaranteed periodicity. The Fourier-based deformation field guarantees loop closure by construction, offering a clear solution to imperfectly looping video supervision.
  • Well-motivated inconsistency handling. The grounded drift field addresses conflicting multi-view supervision, while reference-view anchoring encourages the shared representation to retain meaningful dynamics.
  • Broad empirical evaluation. Experiments on 39 scenes from four sources, complemented by a user study, provide evidence of perceptual benefits over the evaluated baselines.
Critical weaknesses
  • The contribution beyond strong generative priors could be better isolated. The pipeline benefits from powerful pretrained models, particularly GPT-5.5 and Seedance 2.0. Although the ablations provide evidence for the optimization design, several gains are modest: removing the drift field changes orbit-view subject consistency from 0.9673 to 0.9666 and sharpness from 709.75 to 704.34. The larger degradation without scene-view consistent optimization is encouraging, but this component follows Free4D. A comparison with a compatible video-driven baseline using matched reference videos and generation budgets, together with per-scene ablation statistics, would better establish the contribution of the proposed components.

  • The comparison with related methods and broader camera trajectories is incomplete. Make-It-4D[1] is cited but not evaluated. It specifically uses diffusion-based outpainting and layered completion to mitigate missing content under larger camera movements, making it relevant to the disocclusion artifacts highlighted in the comparison with 3D Cinemagraphy. Since other single-image methods are already included, the input difference alone does not adequately explain its exclusion. Comparing against Make-It-4D and testing larger camera translations and rotations would clarify the practical advantages and viewpoint limitations of OuroWorld beyond the current ±20° evaluation.

  • Computational efficiency is insufficiently characterized. The framework involves reference-video generation, VGGT-Ω lifting, multi-view video inpainting, diffusion-based supervision refinement, and per-scene optimization. However, the paper does not provide an end-to-end runtime breakdown or an efficiency comparison with the baselines. Appendix B also states that optimization iterations vary by scene without specifying their range or selection criteria. Reporting generation latency and sampling budgets, optimization time and memory, and final rendering speed would help assess the quality–cost trade-off and scalability to larger scene collections.

  • The novelty of the periodic parameterization needs more precise positioning. Although the paper acknowledges previous Fourier-based dynamic representations, sinusoidal time encoding has also been used explicitly to enforce looping: CycleNet[2] uses integer-frequency sine/cosine encodings to guarantee periodic network outputs. OuroWorld applies this principle to spatially varying Gaussian deformations within a different framework, but the distinction should be discussed more clearly. A comparison with a simpler fixed periodic time encoding under otherwise identical supervision and optimization would help establish the benefit of the proposed parameterization.

[1] Make-it-4d: Synthesizing a consistent long-term dynamic scene video from a single image. ACMMM 2023

[2] Blowing in the Wind: CycleNet for Human Cinemagraphs from Still Images. CVPR 2023

Questions for authors
  • Can you better isolate the contribution of the proposed modules through a compatible baseline using matched video supervision, and report how the ablation gains vary across scenes?
  • Why was Make-It-4D excluded from the experiments, and how does OuroWorld perform under larger camera movements beyond the evaluated ±20° range?
  • What are the per-scene generation and optimization costs, and the final rendering speed? How are optimization iterations selected, and how does efficiency compare with the baselines?
  • How does the periodic parameterization relate to prior cyclic encodings such as CycleNet, and what advantages does it offer over a fixed periodic time encoding under identical training conditions?
4 comments
AuthorsOpenReview ↗

Response to Reviewer P3NJ (Part 1/2)

We thank the reviewer for recognizing (1) a practical framework combining video priors with 3D animation without motion masks; (2) looping by construction via the Fourier deformation; (3) well-motivated handling of inconsistent supervision via the Grounded Drift Field and reference-view anchoring; and (4) a broad 39-scene evaluation with a user study.

W1, Q1 — Contribution beyond the generative priors.

Matched video-driven baseline. We ran Free4D, which also builds 4DGS from generated videos, on all 39 scenes (Table 2 protocol) with our supervision: the same GPT-5.5 prompt and Seedance 2.0 reference video per scene, so the generation budget is identical.

  • Free4D (our ref.): our reference video + Free4D's own multi-view generation and 4D optimization.

  • Free4D (our MV): our reference and multi-view videos; only the 4D stage differs. |||Vividness|Naturalness|Loop Coherence||Scene Quality||||| |-|-|-|-|-|-|-|-|-|-|-| |Camera|Method|Vividness ↑|KVD ↓|MALF ↑|Seam SSIM ↑|Aesthetic Quality ↑|Imaging Quality ↑|Subject Consistency ↑|Background Consistency ↑|Failure Rate ↓| |Static|Free4D (our ref.)|0.0474|70.0|0.0880|0.8996|0.6287|0.6615|0.9873|0.9811|2/39| |Static|Free4D (our MV)|0.0453|65.7|0.0771|0.9665|0.6502|0.6929|0.9915|0.9829|1/39| |Static|Ours|0.0427|61.7|0.0820|0.9963|0.6708|0.7155|0.9921|0.9826|0/39| |Orbit|Free4D (our ref.)|0.4624|57.2|–|–|0.5983|0.5399|0.9498|0.9574|2/39| |Orbit|Free4D (our MV)|0.4833|44.9|–|–|0.6391|0.6524|0.9628|0.9622|1/39| |Orbit|Ours|0.5132|39.1|–|–|0.6615|0.6932|0.9658|0.9610|0/39|

  • Same videos, still better. OuroWorld wins 6/8 static and 5/6 orbit metrics, so the gains come from our 4D representation, not from GPT-5.5 or Seedance 2.0.

  • Loops and moves. Free4D leaves a seam (Seam SSIM 0.900) or loops with less motion (MALF); ours is high on both.

  • Our Stage 2 also helps Free4D. Our MV videos improve its KVD, Seam SSIM, Aesthetic and Imaging.

User study (2AFC as in Fig. 6; % preferring the first method):

ComparisonVisual qualityLoopingVividnessNaturalness
Ours vs. Free4D (our MV)87.4%90.5%88.9%90.0%
Ours vs. Free4D (our ref.)98.9%98.9%98.4%98.9%
Free4D (our MV) vs. Free4D (our ref.)75.6%73.9%71.7%74.4%
  • Same videos, still preferred. Ours wins every criterion; our MV videos alone also make Free4D preferred over its own.

Adding each component to Free4D — Grounded Drift Field (Free4D → + Drift):

VividnessNaturalnessSharpnessScene Quality
SettingCameraVividness ↑KVD ↓VoL ↑Imaging ↑Subject ↑Background ↑
Our ref.Static0.0474 → 0.0518 (+9.3%)70.0 → 59.8 (−14.6%)371.0 → 427.5 (+15.2%)0.6615 → 0.6812 (+3.0%)0.9873 → 0.9847 (−0.3%)0.9811 → 0.9788 (−0.2%)
Our MVStatic0.0453 → 0.0448 (−1.1%)65.7 → 63.4 (−3.6%)804.9 → 832.1 (+3.4%)0.6929 → 0.6935 (+0.1%)0.9915 → 0.9922 (+0.1%)0.9829 → 0.9860 (+0.3%)
Our ref.Orbit0.4624 → 0.4863 (+5.2%)57.2 → 42.8 (−25.2%)195.9 → 256.4 (+30.9%)0.5399 → 0.5855 (+8.5%)0.9498 → 0.9507 (+0.1%)0.9574 → 0.9581 (+0.1%)
Our MVOrbit0.4833 → 0.4911 (+1.6%)44.9 → 37.9 (−15.6%)512.8 → 556.6 (+8.5%)0.6524 → 0.6635 (+1.7%)0.9628 → 0.9643 (+0.2%)0.9622 → 0.9651 (+0.3%)

Periodic Deformation Field (→ + Periodic, static camera):

Loop Coherence
SettingMALF ↑Seam SSIM ↑
Our ref.0.0880 → 0.0913 (+3.8%)0.8996 → 0.9972 (+10.8%)
Our MV0.0771 → 0.0952 (+23.5%)0.9665 → 0.9898 (+2.4%)
  • Drift improves sharpness, naturalness and Imaging in all cases, most with Free4D's own, less consistent MV videos under orbit (VoL +30.9%, KVD −25.2%), as expected from absorbing cross-view disagreement.
  • Periodic deformation closes the seam (Seam SSIM up to 0.997) and raises MALF: looping without freezing motion.
  • Both help a different pipeline, so they are not tied to our setup.
  • Modest gains in Table 3. In our pipeline the Sec. 3.4 objective already handles much of the inconsistency, so the drift fixes local artifacts that scene-level metrics barely register (Fig. 7). With less consistent supervision (Free4D above), its gains are much larger.

W2, Q2 — Make-It-4D. Make-It-4D [1] cannot be run: its repository releases neither the model definition nor checkpoints, and code requests have gone unanswered since 2023. Other single-image methods (3D Cinemagraphy, 3D-MOM) are included because their code runs.

[1] L. Shen et al. Make-It-4D: Synthesizing a Consistent Long-Term Dynamic Scene Video from a Single Image. ACM MM 2023.

(Continued in Part 2: W2/Q2 camera motion, W3/Q3, W4/Q4.)

AuthorsOpenReview ↗

Response to Reviewer P3NJ (Part 2/2)

(Continued from Part 1: W1/Q1, W2/Q2.)

W2, Q2 (cont.) — Larger, diverse camera motion. Without retraining (trained within ±20° yaw), we render all 39 scenes along six extra paths aimed at the orbit pivot: orbit ±30°/±40°; dolly-in (40% closer); spiral-in; figure-8 (adds unseen ±3° pitch); breathing (±6° sway, 2° rise, 8% push-in).

Scene QualitySharpnessNaturalness
Camera pathMax rotationMax translationAesthetic Quality ↑Subject Consistency ↑Background Consistency ↑VoL ↑KVD ↓
Orbit ±20° (paper)20°0.340.6660.9670.962706.450.1
Orbit ±30°30°0.510.658 (−1.2%)0.954 (−1.3%)0.956 (−0.6%)680.2 (−3.7%)53.6 (+7.0%)
Orbit ±40°40°0.670.645 (−3.2%)0.939 (−3.0%)0.949 (−1.3%)661.0 (−6.4%)62.2 (+24.0%)
Dolly-in0°0.400.661 (−0.7%)0.963 (−0.4%)0.964 (+0.1%)764.7 (+8.3%)29.8 (−40.6%)
Spiral-in5.8°0.200.665 (−0.1%)0.970 (+0.3%)0.966 (+0.4%)760.2 (+7.6%)36.8 (−26.5%)
Figure-88°0.140.668 (+0.3%)0.973 (+0.6%)0.967 (+0.5%)740.5 (+4.8%)45.4 (−9.3%)
Breathing6.1°0.110.668 (+0.3%)0.977 (+1.0%)0.966 (+0.4%)750.5 (+6.2%)41.7 (−16.8%)

*Translation relative to camera-to-pivot distance.

  • Wider rotation. At ±30° (1.5× training), scene quality drops ≤1.3%.
  • Unseen axes. Dolly-in moves farther than the orbit (0.40 vs. 0.34), figure-8 adds pitch; these and the combined paths lose ≤0.7% and all improve VoL and KVD.
  • ±40° is the only clear drop: in enclosed scenes (e.g., rooms) the camera exits through walls. We will add this table and state the tested range by "any viewpoint".

W3, Q3 — Computational cost. One RTX 4090. Gaussians-to-Life and LoopGaussian exceed 24 GiB on some scenes, so they ran on an A100 with runtime scaled by the measured A100/4090 ratio.

StageTimePeak VRAM
Dynamics prompt (GPT-5.5)API call—
Reference video (Seedance 2.0)~2.5 min (API call)—
VGGT-Ω lifting~15 s7.8 GiB
TrajectoryCrafter inpainting~3.5 min / view, 70 min total15 GiB
IRP-4DGS optimization1–4 hr9 GiB
End-to-end2.2–5.2 hr15 GiB
Rendering10 FPS3 GiB
MethodTime per scenePeak VRAM (train)Rendering FPSPeak VRAM (render)
Gaussians-to-Life19.7–34.2 min22.8–64.3 GiB2.3–2.91.5–5.7 GiB
3D Cinemagraphy14–37 min4 GiB0.74 GiB
LoopGaussian28.9–59.4 min8.5–35.8 GiB1.8–2.79.9–11.6 GiB
3D-MOM65.7–70.2 min9.2 GiB2.3–2.81.3–2.4 GiB
Ours2.2–5.2 hr15 GiB102.8–3.4 GiB
  • Iterations. 8,000, then +5,000 while visibly unconverged, up to 20,000 (a VLM judge can automate this): 16.7% stop at 8,000, 77.8% within 13,000, 5.6% at 20,000. Baselines use official settings, also checked.
  • Sampling budget. Reference videos are regenerated until GPT-5.5 judges the motion natural, as for all baselines (Sec. 4.2).
  • Trade-off. Slower to build, but offline and once; the 20 inpainting runs parallelize. Rendering is fastest (10 FPS, 3.4× the best baseline), and training peaks at a fixed 15 GiB vs. up to 64.3 GiB.

W4, Q4 — Positioning against CycleNet. CycleNet [2] uses sin/cos time encoding for 2D human cinemagraphs, and our periodicity is built into a 3D representation that deforms Gaussians. The designs differ in where periodicity enters:

  • CycleNet-style: one sin/cos code of t, shared by all positions, is concatenated with the spatial feature; the MLP must create position-specific amplitude/phase.
  • Ours: each triplane position stores its own Fourier coefficients (Eq. 1), so amplitude/phase vary in space directly. Swapping in CycleNet-style encoding (same K, supervision, optimization, parameter count; 39 scenes):
VividnessNaturalnessLoop CoherenceScene Quality
CameraMethodVividness ↑KVD ↓MALF ↑Seam SSIM ↑Aesthetic Quality ↑Overall Consist. ↑
StaticCycleNet-style0.0387 (−8.4%)109.22 (+0.9%)0.0751 (−7.2%)0.9979 (+0.1%)0.6803 (+0.8%)0.2111 (0.0%)
StaticOurs0.0423108.230.08090.99650.67480.2111
OrbitCycleNet-style0.5111 (−0.1%)51.00 (+1.8%)––0.6664 (+0.1%)0.2181 (−0.2%)
OrbitOurs0.511750.11––0.66590.2185

*Percentages relative to Ours.

  • Both loop (Seam SSIM ≥ 0.996).
  • Ours keeps more motion. Static camera: CycleNet-style loses 8.4% Vividness and 7.2% MALF; other metrics stay within 1% (2% under orbit, where camera motion dominates).
  • Why. A shared time code needs a learned space–time interaction; under inconsistent supervision it is easier to ignore time and average views toward a static scene. Per-position coefficients give each region's amplitude a direct gradient. We will add this comparison.

[2] H. Bertiche et al. Blowing in the Wind: CycleNet for Human Cinemagraphs From Still Images. CVPR 2023.

We thank the reviewer again, and are glad to answer further within the remaining discussion time.

Sincerely, The Authors

Thank you very much to the author for responding to my comments; my concerns have been addressed. Taking both the paper and the author's response into account, I maintain my "Weak acceptance" recommendation.

We sincerely thank reviewer P3NJ for the time and thoughtful feedback, and are glad our response addressed the concerns.

Reviewer_PhbtOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Recent 3D world models generate photorealistic, explorable scenes, yet these worlds remain frozen in time. This paper aims to bring them alive by turning any static 3D Gaussian Splatting scene into a 3D cinemagraph: a dynamic scene with vivid motion that loops seamlessly from any viewpoint. Existing 3D cinemagraph methods rely on Eulerian flow and user-provided masks, which restricts them to fluid-like motion. Video diffusion models offer much more diverse dynamics, but their outputs are only approximately periodic and are inconsistent across views. OuroWorld addresses this with a mask-free pipeline. A vision-language model infers plausible scene dynamics and prompts a video model to generate a looping reference video, which is then lifted into a dynamic point cloud and inpainted into multi-view videos. The scene is fit with an Inconsistency-Robust Periodic 4DGS. Its Fourier-series deformation field guarantees looping by construction, and a Grounded Drift Field, anchored at the reference view and used only during training, absorbs cross-view inconsistency. The work shows that motion priors from video models can be distilled into a truly 3D, endlessly looping scene that covers general deformation, object movement and changes in lighting, so generated 3D worlds no longer have to stay frozen.

Critical strengths

1. Learning a looping 4D scene from inconsistent generated videos. A Fourier-series deformation with period T makes looping hold by construction. The more important part is the Grounded Drift Field: per-view drift absorbs the disagreement between generated views, and leaving the reference view without drift keeps the drift from absorbing the dynamics themselves (Fig. 4b). This matters because generated multi-view videos are increasingly used to supervise 4D scenes, and this is a simple, well-ablated way to handle their inconsistency.

2. Mask-free animation with motion beyond fluids. Prior 3D cinemagraph methods need motion masks and rely on Eulerian flow, which limits them to fluid-like motion. OuroWorld uses a VLM and a video model to decide what moves and how, which enables deformation, object motion and illumination change from any static 3DGS. The user study supports this, with OuroWorld preferred over all baselines on every criterion. This is consequential because it makes static generated worlds dynamic without manual annotation.

Critical weaknesses

1. The evaluation does not isolate the contribution. Most of the pipeline is borrowed from existing work, and the multi-view generation and optimization stages closely follow Free4D. The new parts are the periodic deformation field and the grounded drift field, so the paper should show how much these two parts improve the results. However, the baselines are either fluid-only Eulerian methods or a method that does not target looping, so it is not surprising that OuroWorld beats them. Free4D, which also builds 4DGS from generated videos, is not included as a baseline, and the closely related LivingWorld* is not cited. The newly proposed metrics also show only small differences. As a result, it remains unclear how much the two new designs actually contribute.

2. Limited viewpoint range and motion bandwidth. The paper claims that the motion stays consistent from any viewpoint. However, both training and evaluation use only a narrow yaw range around a single reference view, so wider viewpoints are not tested. In addition, the Fourier basis uses only a few low frequencies, so it may not be able to represent fast motions such as flickering light, which the paper presents as examples. These two limits directly affect the main claims of 3D consistency and diverse motion, but the paper does not analyze either of them.

  • LivingWorld: Mun, Hyeongju, et al. "LivingWorld: Interactive 4D World Generation with Environmental Dynamics." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.
Questions for authors

On isolating the contribution (Weakness 1)

Could you add Free4D as a baseline, using the same reference video and multi-view videos? Since the pipeline closely follows Free4D, this would be the most direct way to show what OuroWorld adds. Could you also add your components to Free4D one at a time: first the periodic deformation, then the grounded drift? This would show how much each new design contributes on its own.

On viewpoint range and motion (Weakness 2)

Could you evaluate on camera paths wider than the training range (for example, up to ±40° yaw) and report how quality and consistency change as the angle increases? This would clarify how far the “any viewpoint” claim holds. Could you provide an ablation over the number of Fourier frequencies K? In particular, it would help to show whether fast motions such as flickering light are preserved or smoothed out, for example by comparing the motion in the reference video with that in the final rendering.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Phbt (Part 1/2)

We thank the reviewer for recognizing (1) looping by construction via the Fourier deformation; (2) a simple, well-ablated Grounded Drift Field whose drift-free reference view keeps it from absorbing the dynamics (Fig. 4b); (3) mask-free animation beyond fluids; and (4) a user study preferring OuroWorld on every criterion.

W1, Q1 — Isolating the contribution.

Relation to Free4D. Only the diffusion-refined LPIPS loss (Sec. 3.4) follows Free4D; we credit it and do not claim it. Other stages differ:

StageFree4DOuroWorld
Inputsingle image → 4D, no loop3DGS → looping 3D cinemagraph
1. Videoanimates image with given promptVLM infers dynamics; first = last frame = input view
2. Multi-viewno 3D input3DGS renders improve VGGT-Ω depth
3. 4Dgeneric deformable 4DGSPeriodic + Grounded Drift Fields

Free4D baseline, with our videos. Official code, all 39 scenes, Table 2 protocol:

  • Free4D (our ref.): our reference video + Free4D's own multi-view generation and 4D optimization.

  • Free4D (our MV): our reference and multi-view videos; only the 4D stage differs. |||Vividness|Naturalness|Loop Coherence||Scene Quality||||| |-|-|-|-|-|-|-|-|-|-|-| |Camera|Method|Vividness ↑|KVD ↓|MALF ↑|Seam SSIM ↑|Aesthetic Quality ↑|Imaging Quality ↑|Subject Consistency ↑|Background Consistency ↑|Failure Rate↓| |Static|Free4D (our ref.)|0.0474|70.0|0.0880|0.8996|0.6287|0.6615|0.9873|0.9811|2/39| |Static|Free4D (our MV)|0.0453|65.7|0.0771|0.9665|0.6502|0.6929|0.9915|0.9829|1/39| |Static|Ours|0.0427|61.7|0.0820|0.9963|0.6708|0.7155|0.9921|0.9826|0/39| |Orbit|Free4D (our ref.)|0.4624|57.2|–|–|0.5983|0.5399|0.9498|0.9574|2/39| |Orbit|Free4D (our MV)|0.4833|44.9|–|–|0.6391|0.6524|0.9628|0.9622|1/39| |Orbit|Ours|0.5132|39.1|–|–|0.6615|0.6932|0.9658|0.9610|0/39|

  • Same videos, still better. OuroWorld wins 6/8 static and 5/6 orbit metrics, so the gains come from our 4D representation, not the priors.

  • Loops and moves. Free4D leaves a seam (Seam SSIM 0.900) or loops with less motion (MALF); ours is high on both.

  • Our Stage 2 also helps Free4D. Our MV videos improve its KVD, Seam SSIM, Aesthetic and Imaging.

User study (2AFC as in Fig. 6; % preferring the first method):

ComparisonVisual qualityLoopingVividnessNaturalness
Ours vs. Free4D (our MV)87.4%90.5%88.9%90.0%
Ours vs. Free4D (our ref.)98.9%98.9%98.4%98.9%
Free4D (our MV) vs. Free4D (our ref.)75.6%73.9%71.7%74.4%
  • Same videos, still preferred. Ours wins every criterion; our MV videos alone also make Free4D preferred over its own.

Adding each component to Free4D — Grounded Drift Field (Free4D → + Drift):

VividnessNaturalnessSharpnessScene Quality
SettingCameraVividness ↑KVD ↓VoL ↑Imaging Quality↑Subject Consistency ↑Background Consistency ↑
Our ref.Static0.0474 → 0.0518 (+9.3%)70.0 → 59.8 (−14.6%)371.0 → 427.5 (+15.2%)0.6615 → 0.6812 (+3.0%)0.9873 → 0.9847 (−0.3%)0.9811 → 0.9788 (−0.2%)
Our MVStatic0.0453 → 0.0448 (−1.1%)65.7 → 63.4 (−3.6%)804.9 → 832.1 (+3.4%)0.6929 → 0.6935 (+0.1%)0.9915 → 0.9922 (+0.1%)0.9829 → 0.9860 (+0.3%)
Our ref.Orbit0.4624 → 0.4863 (+5.2%)57.2 → 42.8 (−25.2%)195.9 → 256.4 (+30.9%)0.5399 → 0.5855 (+8.5%)0.9498 → 0.9507 (+0.1%)0.9574 → 0.9581 (+0.1%)
Our MVOrbit0.4833 → 0.4911 (+1.6%)44.9 → 37.9 (−15.6%)512.8 → 556.6 (+8.5%)0.6524 → 0.6635 (+1.7%)0.9628 → 0.9643 (+0.2%)0.9622 → 0.9651 (+0.3%)

Periodic Deformation Field (→ + Periodic, static camera):

Loop Coherence
SettingMALF ↑Seam SSIM ↑
Our ref.0.0880 → 0.0913 (+3.8%)0.8996 → 0.9972 (+10.8%)
Our MV0.0771 → 0.0952 (+23.5%)0.9665 → 0.9898 (+2.4%)
  • Drift improves sharpness, naturalness and Imaging in all cases, most with Free4D's own, less consistent MV videos under orbit (VoL +30.9%, KVD −25.2%), as expected from absorbing cross-view disagreement.
  • Periodic deformation closes the seam (Seam SSIM up to 0.997) and raises MALF: looping without freezing motion.
  • Both help a different pipeline, so they are not tied to our setup. We will add these tables.

"The new metrics show small differences." Agreed for Table 3. In our pipeline the Sec. 3.4 objective already handles much of the inconsistency (removing it costs 22% orbit VoL), so the drift fixes local artifacts that scene-level metrics barely register, visible in Fig. 7 (without it, clouds in Iceland Canyon blur and fade). With less consistent supervision (Free4D above), its gains are much larger.

LivingWorld. Thank you; we will cite and discuss LivingWorld [1].

[1] H. Mun et al. LivingWorld: Interactive 4D World Generation with Environmental Dynamics. ECCV 2026.

(Continued in Part 2: W2/Q2.)

AuthorsOpenReview ↗

Response to Reviewer Phbt (Part 2/2)

(Continued from Part 1: W1/Q1.)

W2, Q2 — Viewpoint range and motion bandwidth.

Wider viewpoints. We rendered all 39 scenes from the existing checkpoints, trained within ±20° yaw (no retraining), along six extra paths under the Table 2 protocol. All cameras look at the same pivot as the paper's orbit:

  • Orbit ±30° / ±40°: the paper's yaw sweep at 1.5× / 2× the trained ±20°.
  • Dolly-in: straight push toward the pivot, ending 40% closer.
  • Spiral-in: a shrinking circle while pushing in.
  • Figure-8: an Infinity-symbol-shaped path that adds ±3° pitch, which training never sees.
  • Breathing: ±6° sway, plus a 2° rise and 8% push-in once per motion cycle. | | | | Scene Quality | | | Sharpness | Naturalness | | --- | --- | --- | --- | --- | --- | --- | --- | | Camera path | Max rotation | Max translation | Aesthetic Quality↑ | Subject Consistency↑ | Background Consistency↑ | VoL ↑ | KVD ↓ | | Orbit ±20° (paper) | 20° | 0.34 | 0.666 | 0.967 | 0.962 | 706.4 | 50.1 | | Orbit ±30° | 30° | 0.51 | 0.658 (−1.2%) | 0.954 (−1.3%) | 0.956 (−0.6%) | 680.2 (−3.7%) | 53.6 (+7.0%) | | Orbit ±40° | 40° | 0.67 | 0.645 (−3.2%) | 0.939 (−3.0%) | 0.949 (−1.3%) | 661.0 (−6.4%) | 62.2 (+24.0%) | | Dolly-in | 0° | 0.40 | 0.661 (−0.7%) | 0.963 (−0.4%) | 0.964 (+0.1%) | 764.7 (+8.3%) | 29.8 (−40.6%) | | Spiral-in | 5.8° | 0.20 | 0.665 (−0.1%) | 0.970 (+0.3%) | 0.966 (+0.4%) | 760.2 (+7.6%) | 36.8 (−26.5%) | | Figure-8 | 8° | 0.14 | 0.668 (+0.3%) | 0.973 (+0.6%) | 0.967 (+0.5%) | 740.5 (+4.8%) | 45.4 (−9.3%) | | Breathing | 6.1° | 0.11 | 0.668 (+0.3%) | 0.977 (+1.0%) | 0.966 (+0.4%) | 750.5 (+6.2%) | 41.7 (−16.8%) |

*Max translation Relative to the camera-to-pivot distance.

  • Wider rotation. At ±30°, which is 1.5× the training range, Aesthetic, Subject and Background drop by at most 1.3%.
  • Unseen axes. Dolly-in translates farther than the paper's orbit (0.40 vs. 0.34), and Figure-8 adds pitch. These paths and the two combined ones lose at most 0.7% on any metric, and all four improve VoL and KVD.
  • ±40°. This is the only path with a clear drop. At this angle, the camera in many enclosed scenes (e.g., rooms) passes through walls and leaves the generated world, which no method can render. We will add this table to the appendix of paper and state the tested range next to the "any viewpoint" wording.

Motion bandwidth.

The supplementary already contains fast flickering light sources that OuroWorld reproduces, e.g., Marble's Cabin Retreat. Note also that K only limits how fast the input feature f(x, t) changes (Eq. 1), not the final motion. The MLP that decodes this feature is non-linear, so it can turn a smooth input into a sharp change, such as a light switching on and off.

Ablation over number of Fourier frequency terms. We retrain the 5 scenes with the highest Vividness, with K = 2, 4 and 8 frequencies:

VividnessNaturalnessLoop CoherenceScene Quality
CameraKVividness ↑KVD ↓MALF ↑Seam SSIM ↑Aesthetic Quality ↑Overall Consist. ↑
Static20.0845 (−16.6%)220.62 (−4.7%)0.1311 (−8.9%)0.9965 (+0.2%)0.6612 (−0.3%)0.2061 (+0.1%)
Static40.1013231.570.14380.99460.66300.2059
Static80.1081 (+6.7%)254.86 (+10.1%)0.1505 (+4.6%)0.9934 (−0.1%)0.6590 (−0.6%)0.2060 (+0.1%)
Orbit20.5382 (−1.9%)141.21 (+2.0%)––0.6502 (+0.8%)0.2200 (−0.6%)
Orbit40.5486138.48––0.64470.2214
Orbit80.5509 (+0.4%)129.84 (−6.2%)––0.6426 (−0.3%)0.2237 (+1.0%)
  • Too few frequencies damp fast motion. K = 2 loses 16.6% Vividness and 8.9% MALF under the static camera.
  • Returns diminish beyond K = 4. K = 8 adds 6.7% Vividness, less than half the gap from K = 2, but costs 44% more peak VRAM (9 → 13 GiB) and 1.7× the training time. Its static KVD also worsens by 10.1%.

We thank the reviewer again, and are glad to answer further within the remaining discussion time.

Sincerely, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback. OuroWorld turns any static 3DGS into a 3D cinemagraph: a VLM decides what should move, a video generation pipeline realizes it, and an Inconsistency-Robust Periodic 4DGS distills it into a 4D scene that loops from any viewpoint. We are encouraged that the reviewers recognize:

  • Looping guaranteed by construction (Phbt, P3NJ). The Fourier-series Periodic Deformation Field returns exactly to t = 0 at t = T, whatever the optimization outcome, so seamless looping no longer depends on the generated video being periodic.
  • A simple, well-motivated way to learn from inconsistent generated views (Phbt, P3NJ). The Grounded Drift Field absorbs per-view disagreement during training. Anchoring it at the reference view keeps the drift from absorbing the dynamics themselves, which the ablation confirms (Fig. 4b, Table 3).
  • Mask-free, diverse dynamics beyond fluids (Phbt, P3NJ). Prior 3D cinemagraph methods need motion masks and Eulerian flow. OuroWorld needs neither and handles deformation, object motion and illumination change, making any static generated world come alive without manual annotation.
  • Broad evaluation with convincing results (Phbt, P3NJ, Gxtg). On 39 scenes from four sources, OuroWorld is preferred over every baseline on every user-study criterion (70.8%–99.0%), and the generated videos are of satisfying quality.

New in this rebuttal:

AnalysisResultReviewer
Camera paths beyond the training range±30° (1.5× training): ≤1.3% drop in scene quality; unseen push-in / pitch / combined paths: ≤0.7%Phbt, P3NJ
Free4D baseline with our same videosGiven our reference video, or even our multi-view videos, Free4D still trails: OuroWorld wins 6/8 static and 5/6 orbit metrics in the stricter setting; Free4D diverges on 1–2/39 scenes, ours on nonePhbt, P3NJ
Our components added to Free4DDrift: orbit VoL +30.9%, KVD −25.2%, Imaging Quality +8.5%; Periodic: Seam SSIM 0.900 → 0.997Phbt, P3NJ
Computational cost breakdown2.2–5.2 hr end-to-end, 15 GiB peak on one RTX 4090; baselines peak at up to 64.3 GiBP3NJ
Triplane vs. per-Gaussian Fourier coefficients3.3–76× fewer dynamics parameters; peak training VRAM ~9 GiB vs. up to 54 GiBGxtg
3D motion mask instead of mask-freeMasked optimization peels surfaces off moving objects (video + new Fig. 12)Gxtg
Fourier bandwidth K ablationOn the 5 vividest scenes: K = 2 loses 16.6% Vividness; K = 8 adds only 6.7% at +44% VRAM and 1.7× training time; Seam SSIM ≥ 0.993 for all KPhbt
CycleNet-style periodic encodingSame parameter count: both loop (Seam SSIM ≥ 0.996), but the shared time encoding loses 8.4% Vividness and 7.2% MALFP3NJ
Drift-magnitude regularizer instead of groundingRecovers only half of the motion lost without grounding (Vividness −3.6% vs. ours) and shrinks the drift by 98.8%Gxtg

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.