Pith. sign in

REVIEW 4 major objections 5 minor 58 references

Generative 4D Scene Gaussian Splatting with Object View-Synthesis Priors

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read GenMOJO turns a single monocular video of a crowded, occluded scene into a complete 4D reconstruction with 3D Gaussians per object, and uses joint splatting plus object-centric generative priors to produce accurate novel views and point…

desk verdict A competent, well-ablated extension of DreamScene4D whose central 3D claims in occluded regions are plausible but unmeasured. read the letter →

arxiv 2506.12716 v1 pith:MUV7DSB5 submitted 2025-06-15 cs.CV

Wen-Hsuan Chu , Lei Ke , Jianmeng Liu , Mingxiao Huo , Pavel Tokmakov , Katerina Fragkiadaki This is my paper
classification cs.CV
keywords 4Dscenegeneration3DGaussiansplattingpointtrackingscoredistillationsamplingmonocularvideoobject-centricpriorsocclusionreasoningnovelviewsynthesis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

GenMOJO is a test-time optimization method that takes a monocular video of a scene with multiple moving objects and heavy occlusions, and produces a 4D reconstruction: a set of deformable 3D Gaussians per object that can be rendered from new viewpoints and whose points can be tracked through time. The key move is to treat each object separately for generative completion while jointly rendering all objects together for occlusion-aware supervision, so that cross-object occlusions and interactions are modeled instead of being handled independently. If its claims hold, monocular video alone, without multi-view rigs or supervised tracking data, can yield coherent dynamic 3D scenes with accurate point trajectories, which would matter for video understanding, robotics, and world modeling. The paper supports the claim with experiments on DAVIS and a new MOSE-based point-tracking benchmark, where it reports better novel-view quality and lower tracking error than prior optimization-based and even supervised baselines.

What carries the argument

The machinery is a compositional 4D Gaussian representation: per-object canonical deformable 3D Gaussians, a K-plane deformation network with motion bases that predict per-frame deformations, differentiable affine warps that map object-centric coordinates into a shared frame-centric coordinate system, and joint alpha-blended rasterization of all objects. Score distillation sampling with a Zero-1-to-3-style view-conditioned diffusion model supplies appearance and geometry for viewpoints never observed in the input video, while instance-mask rendering and depth-aware scaling keep objects separated and correctly ordered in depth.

What would settle it

Render a monocular video of an object with known ground-truth 3D shape and a camera trajectory that includes unconventional poses (e.g., looking up from below), run GenMOJO, and compare the reconstructed geometry and point tracks against ground truth in regions never visible in the input. If the SDS-prior geometry diverges from the true shape, the method's novel views and tracks will show the error; alternatively, a video with jittery predicted depth should make objects visibly 'jump' along the depth axis, matching the paper's stated limitation.

Watch

Extended reading notes

Core claim

The central claim is that object-wise decomposition plus joint Gaussian splatting plus object-centric score distillation is enough to turn a single moving-camera video into a persistent 4D scene representation that generalizes to unseen views and tracks points through occlusions. Concretely, each foreground object is lifted to a canonical set of 3D Gaussians using rendering losses and an SDS loss from a view-conditioned diffusion model; per-frame differentiable affine transformations place these object-centric Gaussians into a common scene coordinate frame; and a K-plane deformation network predicts per-frame position, rotation, scale, and appearance changes. All objects are then splatted together so that rendering losses, optical flow, depth, and rendered instance masks supervise the whole scene at once, teaching the Gaussians where objects occlude one another. The paper reports that removing joint splatting, removing SDS, or replacing object-level SDS with scene-level SDS each degrades motion accuracy and/or novel-view quality, and that the full model outperforms Consistent4D, DreamGaussian4D, DreamScene4D, and Shape of Motion on their metrics, plus a human preference study.

Load-bearing premise

The load-bearing premise is that the object-centric diffusion prior hallucinates correct appearance and geometry for object parts that are never visible in the input video; if that hallucination is wrong, the 4D reconstruction and the point tracks inherit the error.

Editorial extensions

If this is right

  • Monocular videos of cluttered, occluded scenes can be lifted into 360-degree-viewable 4D reconstructions without multi-camera setups.
  • Point trajectories through long occlusions can be recovered by test-time optimization, competing with supervised point trackers trained on large datasets.
  • Object interactions and depth ordering are captured by joint optimization, avoiding the interpenetration artifacts that occur when objects are optimized independently.
  • Better view-synthesis priors and more stable depth estimators should translate directly into better 4D reconstruction and tracking, since the method is test-time and modular.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The method suggests a general recipe, object-centric generative priors plus joint scene-level rendering, that could extend beyond Gaussian splatting to other neural scene representations, so its impact may outlive the specific representation.
  • The MOSE-PTS dataset, introduced for this paper, shows that supervised trackers degrade sharply under heavy occlusion, which may push the point-tracking community toward test-time optimization or hybrid approaches.
  • Because the diffusion prior only sees cropped objects, the method is likely to inherit the prior's blind spots for unusual camera poses; testing with ground-truth 3D data would quantify how much of the reconstruction is hallucination.
  • A natural extension would be to distill the test-time optimization into a feedforward model, which the authors themselves note as future work, trading per-video quality for speed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes GenMOJO, a test-time optimization method that represents each object in a monocular multi-object video as a set of deformable 3D Gaussians, jointly splats all objects into a common scene-centric coordinate frame, and uses object-centric Score Distillation Sampling (SDS) with Zero-1-to-3 to hallucinate unseen viewpoints. It introduces MOSE-PTS, a new benchmark of manual point tracks on MOSE videos, and reports view-synthesis scores, point-tracking errors, ablations, and a user study, claiming superior 4D reconstruction and accurate 2D and 3D point tracks.

Significance. If the claims are supported, the compositional combination of object-centric diffusion priors with occlusion-aware joint Gaussian splatting is a useful step for monocular multi-object 4D generation, and the MOSE-PTS benchmark with manual annotations is a potentially valuable resource. The paper's strengths include explicit cross-object occlusion modeling through joint rendering, instance-mask supervision to prevent Gaussian drift, a new challenging point-tracking benchmark, and a user study with attention checks. However, the evaluation as it stands does not establish geometric accuracy in unobserved regions or 3D tracking accuracy, so the significance of the central claims is conditional on additional validation.

major comments (4)
  1. [§4.2 and §D.1] The CLIP and LPIPS scores for 'novel views' are computed by comparing rendered images at synthetic (elevation, azimuth) offsets to the input reference frame at the same timestep, not to ground-truth frames at those held-out poses. PSNR is reported only for the reference camera. This protocol measures semantic similarity to a view the model has already seen, not novel-view correctness, so Table 1 does not substantiate the claim of faithful novel views or complete 4D reconstruction. Please evaluate on multi-view or synthetic data with ground truth at the novel poses, or add a geometric consistency metric such as depth/Chamfer distance and clearly separate 'appearance plausibility' from 'view correctness'.
  2. [§4.3 and Table 2] The abstract and introduction claim accurate '2D and 3D point tracks,' but all reported point-tracking metrics (ATE, MTE, A-EPE, M-EPE) are computed on 2D trajectories normalized to 256×256 resolution, and no 3D ground-truth comparison is described. The mechanism by which point tracks are extracted from the deformable Gaussians is also not specified. The claimed 3D tracking accuracy is therefore unsupported by the presented evaluation. Please either report a 3D evaluation or explicitly restrict the claim to 2D projected tracks.
  3. [§3.2, Eq. (1), and §4.5] Unobserved and occluded regions are supervised only by the SDS objective from Zero-1-to-3, which drives rendered images toward the diffusion prior's appearance distribution rather than toward measured geometry. The paper's own limitations section admits that the view-conditioned prior struggles with unconventional camera poses, and Figure 8 shows degenerate textures in unseen viewpoints; yet the method's tracking advantage is attributed to inferring complete 4D object geometry in occluded regions. No experiment validates the geometry of unobserved regions. This is the key gap between 'visually plausible' and 'geometrically accurate.' Please add a direct test of unobserved-region geometry, such as synthetic scenes with ground-truth geometry or held-out real viewpoints from a multi-view dataset.
  4. [§4.1 vs. Supplementary A.1] The main text states that MOSE-PTS includes 20 videos and that the evaluation set consists of 15 DAVIS videos plus 20 MOSE videos, while Supplementary A.1 states 'We annotate 30 videos selected from MOSE.' This discrepancy affects the reproducibility of the benchmark and the interpretability of Tables 1 and 2. Please reconcile the video count and specify the exact subset used in every experiment.
minor comments (5)
  1. [§4.2 and Figure 4] The sentence 'We show qualitative comparisons on MOSE and DAVIS videos between in Figure 4' contains a grammatical error, and the caption does not make clear that Shape of Motion is excluded from the novel-view comparisons.
  2. [§B.2 vs. §3.2] The main text says deformation optimization runs for 40× the number of frames, while Supplementary B.2 says 'We optimize for a 35·T steps'; please correct the inconsistency.
  3. [§2, References] TAP-Vid-DAVIS is inconsistently cited as both [7] and [31] in different places; please standardize the reference.
  4. [§D.2] The user study instructs workers to ignore the background, but the rendered videos include background Gaussians and the method is evaluated on full-scene rendering; please clarify whether background quality was excluded from the perceptual comparison and how the instruction was operationalized.
  5. [General] The paper does not state whether code or the MOSE-PTS annotations will be released; for a test-time optimization method and a new benchmark, public code and annotation release are important for reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the SDS prior is external, point tracks are evaluated against manual annotations, and the one self-citation (DreamScene4D) is a comparison baseline rather than a load-bearing premise.

full rationale

The paper's claimed chain is a test-time optimization, not a formal derivation. Inputs are monocular frames, estimated poses/depth/masks, and an external pretrained diffusion prior (Zero-1-to-3 via SDS). Outputs are deformable Gaussians and point tracks. No equation in the paper defines the output in terms of the evaluation target. The SDS objective in Eq. (1) is an external generative prior; it is not fitted to the point-tracking or novel-view benchmarks. Point tracks are extracted from the optimized Gaussians and evaluated against manually annotated MOSE-PTS ground truth, which is not used in any training or optimization loss. The only notable self-citation is DreamScene4D [5], used as a baseline and as prior work; the paper does not import a uniqueness theorem or ansatz from it, and its central contribution (joint multi-object splatting with object-centric SDS) is independently specified in Section 3.2 and supported by ablations. Evaluation weaknesses, such as comparing novel-view metrics to input frames rather than true novel-view ground truth, and not validating 3D track accuracy directly, are correctness risks rather than circular reductions. Therefore no circular step meeting the required evidentiary standard can be exhibited.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method depends on external pretrained components (Zero-1-to-3, SAM2, DepthCrafter, MegaSAM) whose errors are not modeled. No new physical entities are introduced. Several hand-chosen hyperparameters and unreported loss weights control the optimization, and the new MOSE-PTS dataset is author-created.

free parameters (4)
  • Loss-term weights (Lrgb, Lflow, Ldepth, Lclass, Lreg) = not reported
    The total objective is a weighted sum of rendering, flow, depth, class, and regularization losses, but the exact weights are deferred to the supplementary material and not listed there. This is a hand-chosen set of scalars that materially affects the results.
  • Per-object, per-frame scaling factor k_i = Di/Dj from median depth estimates
    Used to place and scale each object in the frame-centric space, computed from monocular depth predictions rather than from ground-truth geometry. The depth model's errors propagate into object placement.
  • Object-crop scale for object-centric lifting = approximately 65% of image size
    Hand-chosen normalization for the object-centric viewing sphere; affects how the SDS prior sees the object and therefore the hallucinated geometry.
  • Gaussian count cap and pruning thresholds = cap 20,000; opacity < 0.01, scale > 0.05, gradient > 0.5
    Hand-set hyperparameters controlling pruning and densification, chosen partly for memory constraints. They affect final geometry density and quality.
assumptions (4)
  • domain assumption The object-centric generative prior (Zero-1-to-3 with SDS) produces accurate novel-view appearance and geometry for cropped objects.
    Introduced in Section 3.2 where SDS is used to "hallucinate unseen views and complete object geometry"; the paper's own limitations note this prior fails on unconventional camera poses.
  • domain assumption Monocular depth estimates (DepthCrafter/MegaSAM) and mask trackers (SAM2) provide sufficiently accurate initialization and per-frame scaling for object placement.
    Used in Section 3.2 to initialize transformations and depth-aware scaling; the limitations section states that jittery depth can cause objects to jump along the depth axis.
  • domain assumption Deformable 3D Gaussians with K-plane motion bases can represent the dynamics of arbitrarily complex multi-object scenes.
    Adopted in Section 3.2; the representation is inherited from prior work and not independently verified here for heavily occluded scenes.
  • domain assumption Joint alpha-blended Gaussian splatting with per-Gaussian instance labels correctly models cross-object occlusion and prevents object drift.
    Central to the joint rendering step; relies on the differentiable rendering approximation of occlusion and on the instance label loss to keep Gaussians on the correct object.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative 4D Scene Gaussian Splatting with Object View-Synthesis Priors." pith.science (2026). https://pith.science/paper/MUV7DSB5

@misc{pith2026250612716,
  author       = {Pith},
  title        = {Pith review of: Generative 4D Scene Gaussian Splatting with Object View-Synthesis Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MUV7DSB5}},
  note         = {Machine review of arXiv:2506.12716}
}
read the original abstract

We tackle the challenge of generating dynamic 4D scenes from monocular, multi-object videos with heavy occlusions, and introduce GenMOJO, a novel approach that integrates rendering-based deformable 3D Gaussian optimization with generative priors for view synthesis. While existing models perform well on novel view synthesis for isolated objects, they struggle to generalize to complex, cluttered scenes. To address this, GenMOJO decomposes the scene into individual objects, optimizing a differentiable set of deformable Gaussians per object. This object-wise decomposition allows leveraging object-centric diffusion models to infer unobserved regions in novel viewpoints. It performs joint Gaussian splatting to render the full scene, capturing cross-object occlusions, and enabling occlusion-aware supervision. To bridge the gap between object-centric priors and the global frame-centric coordinate system of videos, GenMOJO uses differentiable transformations that align generative and rendering constraints within a unified framework. The resulting model generates 4D object reconstructions over space and time, and produces accurate 2D and 3D point tracks from monocular input. Quantitative evaluations and perceptual human studies confirm that GenMOJO generates more realistic novel views of scenes and produces more accurate point tracks compared to existing approaches.

Figures

Figures reproduced from arXiv: 2506.12716 by the authors.

Figure 1
Figure 1. GenMOJO addresses video-to-4D generation for real-world scenes with many moving objects and heavy occlusions. It creates complete 4D scene reconstructions, supports 360◦ novel view synthesis, and accurately tracks how points move over time. We show examples of rendered images from different viewpoints and time steps, highlighting the motion of a single object for clarity. titative findings. 2. Related work Dynamic S… view at source ↗
Figure 2
Figure 2. Overview of GenMOJO. Given a monocular video, its camera pose trajectory, and segmentation tracks for each dynamic object, our model generates a set of deforming 3D Gaussians per object. This is achieved through object-centric Score Distillation Sampling (SDS) and frame-centric joint Gaussian splatting, which accounts for cross-object occlusions by imposing rendering losses on rendered instance masks, point tracks, … view at source ↗
Figure 3
Figure 3. Sample annotation from MOSE-PTS. We visualize annotated points that are not occluded, with corresponding point tracks displayed in matching colors. Implementation Details. All experiments were con￾ducted on a 48GB NVIDIA A6000 GPU. For object-centric lifting, we crop and scale the individual objects to approxi￾mately 65% of the image size. Static 3D Gaussian optimiza￾tion is performed over 1,000 iterations with a ba… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Video to 4D Scene Generation Comparisons. We render Shape of Motion [46], DreamScene4D [5], and GenMOJO on a DAVIS video (left) and a MOSE video (right) from the reference view (top three rows) and a novel view (bottom two rows). We can see that the baselines produce a…
Figure 5
Figure 5. Figure 5: Statistical comparison between MOSE-PTS and Tap-Vid-DAVIS [7]. Avg. number of points: the labeled points in each dataset. Avg. OCC. Rate (Traj / Video): the mean occlusion rate across individual trajectories, capturing the frequency of occlusions within distinct trajec…
Figure 6
Figure 6. Figure 6: The annotation interface comprises four key components: a visualization panel for reviewing and marking points, button panel [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Tracking comparisons. GenMOJO produces more accurate point tracks compared to other optimization-based methods like Shape of Motion and DreamScene4D. CoTrackerV3 produces very accurate point tracks when it succeeds, but also produces tremendous errors when it fails, as…
Figure 8
Figure 8. Figure 8: Failure cases. Top: Erroneous depth predictions can cause entities to jitter along the depth dimension. Bottom: View synthesis diffusion model failures result in degenerate textures in unseen viewpoints. To judge the quality of the videos, consider the following points…
Figure 9
Figure 9. Figure 9: User survey interface. A GUI sample of what an Amazon Turk worker sees for the user study. 5 [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 26 canonical work pages

  1. [1]

    4d-fy: Text-to-4d generation using hybrid score distillation sampling.arXiv preprint arXiv:2311.17984, 2023

    Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling.arXiv preprint arXiv:2311.17984, 2023. 2

    arXiv 2023

  2. [2]

    TC4D: Trajectory-Conditioned Text-to-4D Generation

    Sherwin Bahmani, Xian Liu, Yifan Wang, Ivan Sko- rokhodov, Victor Rong, Ziwei Liu, Xihui Liu, Jeong Joon Park, Sergey Tulyakov, Gordon Wetzstein, et al. Tc4d: Trajectory-conditioned text-to-4d generation.arXiv preprint arXiv:2403.17920, 2024. 2

    work page Pith review arXiv 2024

  3. [3]

    FlowIBR: Leveraging Pre-Training for Efficient Neural Image-Based Rendering of Dynamic Scenes

    Marcel B ¨usching, Josef Bengtson, David Nilsson, and M˚arten Bj¨orkman. Flowibr: Leveraging pre-training for effi- 9 cient neural image-based rendering of dynamic scenes.arXiv preprint arXiv:2309.05418, 2023. 2

    work page Pith review arXiv 2023

  4. [4]

    Hexplane: A fast representa- tion for dynamic scenes

    Ang Cao and Justin Johnson. Hexplane: A fast representa- tion for dynamic scenes. InCVPR, 2023. 2, 5

    work page 2023

  5. [5]

    Dream- scene4d: Dynamic multi-object scene generation from monocular videos

    Wen-Hsuan Chu, Lei Ke, and Katerina Fragkiadaki. Dream- scene4d: Dynamic multi-object scene generation from monocular videos. InNeurIPS, 2024. 1, 3, 5, 6, 7, 8, 9, 2

    work page 2024

  6. [6]

    Mose: A new dataset for video object segmentation in complex scenes

    Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. Mose: A new dataset for video object segmentation in complex scenes. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 20224–20234, 2023. 1, 3, 6

    work page 2023

  7. [7]

    Tap-vid: A benchmark for tracking any point in a video

    Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Re- casens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. InNeurIPS, 2022. 5, 6, 7, 8, 1

    work page 2022

  8. [8]

    K-planes: Explicit radiance fields in space, time, and appearance

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2

    2023

Show all 58 references
  1. [9]

    Gaussianflow: Splatting gaussian dynamics for 4d content creation.arXiv preprint arXiv:2403.12365, 2024

    Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wen- chao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splatting gaussian dynamics for 4d content creation.arXiv preprint arXiv:2403.12365, 2024. 2

    2024 arXiv

  2. [10]

    Kubric: A scalable dataset generator

    Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapra- gasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. InCVPR, 2022. 8, 9

    2022

  3. [11]

    Particle video revisited: Tracking through occlusions using point trajectories

    Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. InEuropean Conference on Computer Vi- sion, pages 59–75. Springer, 2022. 8

    2022

  4. [12]

    Inferring articulated rigid body dynam- ics from rgbd video

    Eric Heiden, Ziang Liu, Vibhav Vineet, Erwin Coumans, and Gaurav S Sukhatme. Inferring articulated rigid body dynam- ics from rgbd video. InIROS, 2022. 1

    2022

  5. [13]

    Depthcrafter: Generating consistent long depth sequences for open-world videos.arXiv preprint arXiv:2409.02095, 2024

    Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos.arXiv preprint arXiv:2409.02095, 2024. 3, 4

    2024 arXiv

  6. [14]

    Consistent4d: Consistent 360{\deg}dynamic object gen- eration from monocular video

    Yanqin Jiang, Li Zhang, Jin Gao, Weimin Hu, and Yao Yao. Consistent4d: Consistent 360{\deg}dynamic object gen- eration from monocular video. InICLR, 2024. 1, 2, 5, 7, 8

    2024

  7. [15]

    Co- tracker: It is better to track together.arXiv preprint arXiv:2307.07635, 2023

    Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co- tracker: It is better to track together.arXiv preprint arXiv:2307.07635, 2023. 3, 5, 8, 1

    2023 arXiv

  8. [16]

    Co- tracker3: Simpler and better point tracking by pseudo- labelling real videos.arXiv preprint arXiv:2410.11831,

    Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co- tracker3: Simpler and better point tracking by pseudo- labelling real videos.arXiv preprint arXiv:2410.11831,

    arXiv

  9. [17]

    Segment anything in high quality

    Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. InNeurIPS, 2023. 3

    2023

  10. [18]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4):1–14, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4):1–14, 2023. 2, 3, 5

    2023

  11. [19]

    Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds.arXiv preprint arXiv:2405.17421, 2024

    Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds.arXiv preprint arXiv:2405.17421, 2024. 2

    2024 arXiv

  12. [20]

    Neural 3d video synthesis from multi-view video

    Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In CVPR, 2022. 2

    2022

  13. [21]

    Megasam: Accurate, fast and ro- bust structure and motion from casual dynamic videos

    Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holyn- ski, and Noah Snavely. Megasam: Accurate, fast and ro- bust structure and motion from casual dynamic videos. In Proceedings of the Computer Vision and Pattern Re...

    2025

  14. [22]

    Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models

    Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fi- dler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763, 2023. 2

    2023 arXiv

  15. [23]

    Neural sparse voxel fields.Advances in Neural Information Processing Systems, 33:15651–15663,

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields.Advances in Neural Information Processing Systems, 33:15651–15663,

  16. [24]

    Zero-1-to-3: Zero-shot one image to 3d object

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tok- makov, Sergey Zakharov, and Carl V ondrick. Zero-1-to-3: Zero-shot one image to 3d object. InICCV, 2023. 1, 2, 3, 5

    2023

  17. [25]

    Robust dynamic radiance fields

    Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Jo- hannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. InCVPR, 2023. 2

    2023

  18. [26]

    Neural vol- umes: Learning dynamic renderable volumes from images

    Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural vol- umes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2

    1906 arXiv

  19. [27]

    Dynamic 3d gaussians: Tracking by per- sistent dynamic view synthesis

    Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by per- sistent dynamic view synthesis. In3DV, 2024. 2, 3, 5

    2024

  20. [28]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. InECCV, 2020. 2

    2020

  21. [29]

    Fast dy- namic 3d object generation from a single-view video.arXiv preprint arXiv 2401.08742, 2024

    Zijie Pan, Zeyu Yang, Xiatian Zhu, and Li Zhang. Fast dy- namic 3d object generation from a single-view video.arXiv preprint arXiv 2401.08742, 2024. 2

    2024

  22. [30]

    Nerfies: Deformable neural radiance fields

    Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. InICCV, 2021. 2

    2021

  23. [31]

    The 2017 davis challenge on video object segmentation.arXiv preprint arXiv:1704.00675, 2017

    Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Ar- bel´aez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation.arXiv preprint arXiv:1704.00675, 2017. 5, 8 10

    2017 arXiv

  24. [32]

    Dreamfusion: Text-to-3d using 2d diffusion.ICLR,

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion.ICLR,

  25. [33]

    D-nerf: Neural radiance fields for dynamic scenes

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. InCVPR, 2021. 2

    2021

  26. [34]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InICML, 2021. 7

    2021

  27. [35]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024. 3, 4

    2024 arXiv

  28. [36]

    Dreamgaussian4d: Genera- tive 4d gaussian splatting.arXiv preprint arXiv:2312.17142,

    Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Genera- tive 4d gaussian splatting.arXiv preprint arXiv:2312.17142,

    arXiv

  29. [37]

    Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159,

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159,

    arXiv

  30. [38]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InCVPR, 2022. 5

    2022

  31. [39]

    Zeronvs: Zero-shot 360- degree view synthesis from a single real image

    Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry La- gun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360- degree view synthesis from a single real image. InCVPR,

  32. [40]

    Dynomo: Online point tracking by dynamic online monocular gaussian recon- struction.arXiv preprint arXiv:2409.02104, 2024

    Jenny Seidenschwarz, Qunjie Zhou, Bardienus Duisterhof, Deva Ramanan, and Laura Leal-Taix ´e. Dynomo: Online point tracking by dynamic online monocular gaussian recon- struction.arXiv preprint arXiv:2409.02104, 2024. 3

    2024 arXiv

  33. [41]

    Self-supervised any- point tracking by contrastive random walks.ECCV, 2024

    Ayush Shrivastava and Andrew Owens. Self-supervised any- point tracking by contrastive random walks.ECCV, 2024. 8

    2024

  34. [42]

    Text-to-4d dy- namic scene generation.arXiv preprint arXiv:2301.11280,

    Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dy- namic scene generation.arXiv preprint arXiv:2301.11280,

    arXiv

  35. [43]

    Dynamic gaussian marbles for novel view synthesis of casual monocular videos.arXiv preprint arXiv:2406.18717, 2024

    Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos.arXiv preprint arXiv:2406.18717, 2024. 2

    2024 arXiv

  36. [44]

    Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation.ICLR, 2024

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation.ICLR, 2024. 3, 1, 2

    2024

  37. [45]

    Tracking everything everywhere all at once

    Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 19795–19806, 2023. 3

    2023

  38. [46]

    Shape of motion: 4d reconstruc- tion from a single video.arXiv preprint arXiv:2407.13764,

    Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruc- tion from a single video.arXiv preprint arXiv:2407.13764,

  39. [47]

    Gflow: Recovering 4d world from monocular video.arXiv preprint arXiv:2405.18426, 2024

    Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video.arXiv preprint arXiv:2405.18426, 2024. 2

    2024 arXiv

  40. [48]

    4d gaussian splatting for real-time dynamic scene rendering

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. InCVPR, 2024. 2, 3

    2024

  41. [49]

    Physgaussian: Physics- integrated 3d gaussians for generative dynamics

    Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics- integrated 3d gaussians for generative dynamics. InCVPR,

  42. [50]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. InCVPR, 2024. 3

    2024

  43. [51]

    Depth any- thing v2.arXiv:2406.09414, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2.arXiv:2406.09414, 2024. 9

    2024 arXiv

  44. [52]

    Beyond skeletons: Integrative latent mapping for coherent 4d sequence genera- tion.arXiv preprint arXiv:2403.13238, 2024

    Qitong Yang, Mingtao Feng, Zijie Wu, Shijie Sun, Weisheng Dong, Yaonan Wang, and Ajmal Mian. Beyond skeletons: Integrative latent mapping for coherent 4d sequence genera- tion.arXiv preprint arXiv:2403.13238, 2024. 2

    2024 arXiv

  45. [53]

    Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction.arXiv preprint arXiv:2309.13101, 2023

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction.arXiv preprint arXiv:2309.13101, 2023. 2

    2023 arXiv

  46. [54]

    4dgen: Grounded 4d content gen- eration with spatial-temporal consistency.arXiv preprint arXiv:2312.17225, 2023

    Yuyang Yin, Dejia Xu, Zhangyang Wang, Yao Zhao, and Yunchao Wei. 4dgen: Grounded 4d content gen- eration with spatial-temporal consistency.arXiv preprint arXiv:2312.17225, 2023. 1, 2

    2023 arXiv

  47. [55]

    Stag4d: Spatial-temporal anchored generative 4d gaussians.arXiv preprint arXiv:2403.14939, 2024

    Yifei Zeng, Yanqin Jiang, Siyu Zhu, Yuanxun Lu, Youtian Lin, Hao Zhu, Weiming Hu, Xun Cao, and Yao Yao. Stag4d: Spatial-temporal anchored generative 4d gaussians.arXiv preprint arXiv:2403.14939, 2024. 1, 2

    2024 arXiv

  48. [56]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InCVPR, 2018. 7

    2018

  49. [57]

    Pointodyssey: A large-scale synthetic dataset for long-term point tracking

    Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wet- zstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. InICCV,

  50. [58]

    occluded

    Yufeng Zheng, Xueting Li, Koki Nagano, Sifei Liu, Otmar Hilliges, and Shalini De Mello. A unified approach for text-and image-guided 4d scene generation.arXiv preprint arXiv:2311.16854, 2023. 2 11 Generative 4D Scene Gaussian Splatting with Object View-Synthesis Priors Supplem...

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.