Pith. sign in

REVIEW 3 major objections 5 minor 64 references

ViDAR: Video Diffusion-Aware 4D Reconstruction From Monocular Inputs

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ViDAR shows that diffusion-enriched pseudo-views, masked to dynamic regions and aligned by pose optimisation, improve monocular 4D reconstruction beyond prior methods on DyCheck.

desk verdict A solid empirical extension of diffusion-guided supervision to dynamic monocular reconstruction, with an honest limitations section but an over-strong headline claim and a few methodological soft spots. read the letter →

arxiv 2506.18792 v1 pith:VYZOVV3K submitted 2025-06-23 cs.CV

Michal Nazarczuk , Sibi Catley-Chandar , Thomas Tanay , Zhensong Zhang , Gregory Slabaugh , Eduardo Pérez-Pellitero This is my paper
classification cs.CV
keywords dynamicnovelviewsynthesismonocularvideo4Dreconstructiondiffusionmodels3DGaussiansplattingpseudomulti-viewsupervisioncameraposeoptimisationDyCheckbenchmarkspatio-temporalconsistency
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

ViDAR is a method for turning a single casually captured video of a dynamic scene into a 4D reconstruction that can be re-rendered from new viewpoints. The paper argues that a diffusion model personalised to the scene can act as a useful source of pseudo multi-view supervision, provided its outputs are used carefully: supervision is applied only to dynamic regions, a perceptual loss absorbs texture flicker, and the sampled camera poses are optimised to keep the synthetic views aligned with real geometry. The claim is that this diffusion-aware scheme yields reconstructions that beat existing monocular 4D methods on the DyCheck benchmark in visual quality and geometric consistency, with the largest margins on moving parts of the scene. The paper also contributes a dynamic-region evaluation protocol, since standard co-visibility masks are dominated by static pixels.

What carries the argument

The central object is the pseudo multi-view training set $\{E_{m,t}\}$ produced by rendering novel cameras from an initial monocular reconstruction and then enhancing each render with a scene-personalised Stable Diffusion XL model using a k-step noising and denoising image-to-image process. The design that carries the argument is the diffusion-aware loss: $L_{\mathrm{dyn}}$ supervises only dynamic regions with L1 plus perceptual and SSIM terms, while $L_{\mathrm{cam}}$ supervises the full image to optimise the sampled camera poses, and the two are backpropagated in separate passes so that pose alignment and geometry and texture refinement do not interfere.

What would settle it

Run ViDAR exactly as described but replace the personalised diffusion enhancement with the identity, using the original degraded renders as pseudo ground truth while keeping the dynamic masking, perceptual losses, and camera-pose optimisation. If co-visibility PSNR-m and dynamic PSNR-D do not drop by roughly the margins reported in the paper's tables, the diffusion enhancement is not what drives the gains. A second check is to record the optimised sampled camera poses: if they drift far beyond the noise level used for sampling, the pose optimisation is absorbing synthetic geometry rather than aligning views.

Watch

Extended reading notes

Core claim

The central claim is that the main obstacle to using diffusion-generated views for monocular 4D reconstruction is not their visual quality but their spatio-temporal inconsistency, and that this obstacle can be overcome by a diffusion-aware training scheme rather than by better diffusion models. The pipeline begins with a MoSca-style Gaussian splatting reconstruction, renders novel views from sampled cameras, enhances those renders with a DreamBooth-personalised SDXL model, and then retrains the reconstruction using the enhanced images as pseudo ground truth. The two mechanisms that make this work are a dynamic-region mask that restricts the pseudo-view loss to moving content, and a camera-pose optimisation that aligns the sampled cameras to the scene geometry using the coarse static structure in the full enhanced images. On DyCheck the resulting reconstructions outperform prior monocular methods on co-visibility-masked metrics and by a wider margin on the newly introduced dynamic-mask metrics.

Load-bearing premise

The load-bearing premise is that diffusion-enhanced pseudo-views, which the paper itself notes are spatio-temporally inconsistent and can contain hallucinated geometry in unseen viewpoints, are still net-positive supervision once masked to dynamic regions and aligned by camera-pose optimisation.

Editorial extensions

If this is right

  • Dynamic regions, the hardest part of monocular 4D reconstruction, stand to gain the most from diffusion-based pseudo supervision: ViDAR reports its largest relative improvements on dynamic-mask metrics.
  • Because the pseudo-views are generated offline once and reused, the additional training cost is mostly in the increased refinement iterations, making the approach practical on a single GPU.
  • The dynamic-region benchmark with Track Anything masks can be adopted by future monocular 4D methods to report performance where current co-visibility metrics are uninformative.
  • Naively supervising with diffused views degrades reconstruction, so any method that wants to use diffusion priors for 4D needs an explicit consistency-handling mechanism; the ablations show that removing the camera-pose or dynamic-mask components drops reconstruction quality.

Reading between the lines

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

  • The same diffusion-aware recipe could transfer to other reconstruction backbones: nothing in the argument is specific to MoSca, so a stronger or faster monocular baseline should compound the gains.
  • The camera-pose optimisation creates a risk of a confound: if the sampled poses drift to fit synthetic texture rather than true geometry, reported gains could partly come from overfitting to the diffusion model's hallucinated structure; a test with identity enhancement would separate the two effects.
  • The dynamic-mask protocol could generalise beyond DyCheck to any casually captured monocular video, and might be combined with uncertainty weighting of pseudo-views to down-weight regions where the diffusion model is most likely to hallucinate.
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

3 major / 5 minor

Summary. The paper introduces ViDAR, a pipeline for monocular 4D reconstruction that first builds an initial reconstruction with MoSca, then samples novel camera trajectories, enhances the corresponding renders with a DreamBooth-personalized SDXL diffusion model, and finally retrains the 4D representation using a diffusion-aware loss. The diffusion-aware loss applies supervision only to dynamic regions (masked by Track Anything) and includes a separate camera-pose optimization step for the sampled views. The method is evaluated on the DyCheck benchmark under the standard co-visibility metrics, a newly proposed dynamic-region metric, and ablations. The authors claim state-of-the-art performance in visual quality and geometric consistency, with particular gains in dynamic regions.

Significance. If the claims are validated, the work makes a useful contribution: it demonstrates a plausible alternative to SDS-style diffusion guidance for monocular dynamic reconstruction, and the proposed dynamic-region benchmark addresses a real weakness of co-visibility masking, which heavily weights static content. The writing is clear, the ablations isolate the main components, and the qualitative comparisons are informative. However, the central claim that diffusion-generated pseudo-views provide net-positive supervision is not independently verified, the main comparison to MoSca is not compute-matched, and the dynamic-region evaluation shares the same masks used for supervision. These issues are load-bearing for the headline conclusion and require additional evidence before the comparison can be considered conclusive.

major comments (3)
  1. [Sec. 3.2.1, Sec. 3.3, Limitations] The pseudo-multi-view supervision is generated by enhancing renders of the authors' own initial reconstruction (Sec. 3.2.1) and is never validated against real scene geometry. The paper itself concedes in Sec. 3.3 that coarse geometry may be hallucinated in novel viewpoints, and the Limitations state that diffusion cannot repair major geometrical artefacts. Since Ldyn supervises dynamic regions with these pseudo-views and Lcam optimizes sampled camera poses against the full pseudo-views, the reported DyCheck gains over MoSca (Tables 1-2) could reflect the diffusion prior's plausible appearance rather than a more accurate 4D model. A concrete test would be to report multi-view reprojection consistency of the enhanced pseudo-views across the M sampled cameras, and to compare the optimized sampled camera poses against the known DyCheck camera poses for held-out views. Without such a check, the central claim that the pseudo-views are net-positive supervision is not established.
  2. [Supp. B.5, Tables 1-2] ViDAR is trained for 40,000 iterations whereas the MoSca baseline is trained for 8,000 iterations (Supp. B.5), and each ViDAR iteration contains two separate forward/backward passes. The comparison is therefore not compute-matched, so the reported improvements over MoSca (+0.56 dB PSNR-m and +0.83 dB PSNR-D in Tables 1-2) may in part reflect longer optimization rather than the proposed diffusion-aware mechanism. Please provide a compute-matched MoSca baseline (e.g., MoSca trained for 40,000 iterations) or an ablation at matched iteration counts, and report wall-clock time or a FLOP-equivalent measure.
  3. [Sec. 4.2.1, Sec. 3.3.1, Table 2] The new dynamic-region benchmark uses the same Track Anything masks that are used inside the method to supervise dynamic regions (Sec. 3.3.1), so Table 2 does not independently validate the dynamic-region improvement. The reported dynamic-region gains could be inflated because the model is explicitly trained to agree with those masks and the evaluation is computed on the same masks. Please validate the masks with an external source (e.g., human annotation on a subset or a different tracker), and report dynamic-region metrics on a mask set not used during training.
minor comments (5)
  1. [Sec. 4.3, Supp. A.2, Tables 1-5] The main-text claim that ViDAR 'outperforms all' baselines is qualified by the per-scene results: for Apple co-visibility, half-resolution PSNR-m is 19.18 versus MoSca's 19.40, and for Paper, LPIPS-D is 0.3080 versus MoSca's 0.2832. Please soften the abstract and Sec. 4.3 claims to 'on average' and explicitly discuss these exceptions.
  2. [Sec. 3.3.1] There is a typo, 'psuedo-GTs', and the loss definitions Ldyn and Lcam are not numbered, which makes it unnecessarily hard to refer to them in the discussion.
  3. [Tables 1, 2, 4] No error bars, confidence intervals, or multi-seed results are reported; given the stochastic diffusion enhancement and camera sampling, the authors should report variance across at least three seeds or explicitly state that results are single-seed.
  4. [Sec. 3.2.1, Supp. B.4] The main text defines M sampled cameras but does not give its value; only Supp. B.4 reveals that 18 cameras per timestep are used. Please state this value in the main text.
  5. [Reproducibility] No code, checkpoints, or released masks are mentioned. For a benchmark contribution, releasing the dynamic masks, sampled camera trajectories, and training code would greatly improve reproducibility and uptake.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: ViDAR's reported gains are evaluated on held-out real test views, and no load-bearing argument reduces to its own inputs or to a self-citation chain.

full rationale

ViDAR is an empirical pipeline rather than a formal derivation, and its central claims are benchmark comparisons against held-out DyCheck test views with real ground truth, not against its own pseudo-views. The pseudo-multi-view supervision is indeed produced from the initial MoSca reconstruction and enhanced by a per-scene DreamBooth-adapted SDXL model (Sec. 3.2, 3.2.1), and the final representation is retrained with this self-generated data (Sec. 3.3). This is a self-training loop, and the paper explicitly concedes that diffusion enhancement cannot repair major geometric artefacts and that coarse geometry may be hallucinated in novel viewpoints (Sec. 3.3; Limitations). Those are correctness risks, not circularity: the final metrics are computed on real test frames, so any hallucination reinforced during retraining would be penalized by the benchmark rather than guaranteed to inflate scores. The dynamic-region benchmark (Table 2) uses Track Anything masks, the same external segmentation tool used inside the method; this aligns the evaluation region with the training objective but still scores actual test-image pixels, so it is a potential confound, not an equation-level reduction. All load-bearing components cited for the method itself, including MoSca, DreamBooth, SDXL, Track Anything, and ReconFusion, are external works; the authors' own SWinGS citation appears only in related work and is not load-bearing. I find no specific step where a claimed prediction or first-principles result is equal, by construction or by self-citation, to its inputs.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The free parameters are design choices for the pipeline, most notably the unreported diffusion step count k. The assumptions concern the reliability of the learned components, MoSca, SDXL, and Track Anything, and the net benefit of synthetic supervision.

free parameters (5)
  • lambda_p, lambda_s (loss weights) = 0.1
    Chosen by hand for the dynamic and camera losses; affects the balance between pixel and perceptual terms.
  • k (diffusion denoising steps in enhancement) = not reported
    Number of noise and denoising steps in the image enhancement process; central to the method but never specified.
  • M (sampled cameras per timestep) = 18
    Number of new camera poses generated per timestamp in the sampling strategy detailed in supplementary Section B.3.
  • Training iterations = 40000
    Increased from MoSca's 8000 to accommodate the additional generated data; this also changes the compute budget versus the baseline.
  • DreamBooth LoRA training steps = 5000
    Set in supplementary Section B.2 to adapt SDXL to more than 400 input frames.
assumptions (5)
  • domain assumption MoSca provides a sufficiently accurate initial reconstruction of the scene.
    The entire pipeline builds on this reconstruction; the limitation section states that ViDAR cannot repair major geometrical artefacts.
  • ad hoc to paper The personalized SDXL model preserves coarse geometry while enhancing fine appearance.
    This assumption underpins the use of diffused images as pseudo-multi-view supervision, and is stated implicitly in Sec. 3.2.
  • domain assumption Track Anything dynamic masks accurately separate dynamic and static regions.
    Masks are used both to restrict diffusion supervision to dynamic regions and to define the new evaluation benchmark.
  • ad hoc to paper Sampled camera poses are close enough to the true geometry that the camera loss can align them.
    The method optimizes sampled poses, which are interpolated from noisy input poses, relying on static-region structure to provide a stable signal.
  • ad hoc to paper Diffusion-generated images, despite inconsistency, provide net-positive training signal after dynamic masking.
    This is the core assumption of the diffusion-aware reconstruction stage and is directly challenged by the paper's own admission of possible hallucinated geometry.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ViDAR: Video Diffusion-Aware 4D Reconstruction From Monocular Inputs." pith.science (2026). https://pith.science/paper/VYZOVV3K

@misc{pith2026250618792,
  author       = {Pith},
  title        = {Pith review of: ViDAR: Video Diffusion-Aware 4D Reconstruction From Monocular Inputs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VYZOVV3K}},
  note         = {Machine review of arXiv:2506.18792}
}
read the original abstract

Dynamic Novel View Synthesis aims to generate photorealistic views of moving subjects from arbitrary viewpoints. This task is particularly challenging when relying on monocular video, where disentangling structure from motion is ill-posed and supervision is scarce. We introduce Video Diffusion-Aware Reconstruction (ViDAR), a novel 4D reconstruction framework that leverages personalised diffusion models to synthesise a pseudo multi-view supervision signal for training a Gaussian splatting representation. By conditioning on scene-specific features, ViDAR recovers fine-grained appearance details while mitigating artefacts introduced by monocular ambiguity. To address the spatio-temporal inconsistency of diffusion-based supervision, we propose a diffusion-aware loss function and a camera pose optimisation strategy that aligns synthetic views with the underlying scene geometry. Experiments on DyCheck, a challenging benchmark with extreme viewpoint variation, show that ViDAR outperforms all state-of-the-art baselines in visual quality and geometric consistency. We further highlight ViDAR's strong improvement over baselines on dynamic regions and provide a new benchmark to compare performance in reconstructing motion-rich parts of the scene. Project page: https://vidar-4d.github.io

Figures

Figures reproduced from arXiv: 2506.18792 by the authors.

Figure 1
Figure 1. ViDAR provides a novel framework for Monocular Novel View Synthesis utilising a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. A high-level overview of ViDAR. The input video is used to create a 4D reconstruction with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An example of co-visibility and dynamic mask comparison. regions, the dynamic regions are arguably the area of most interest and also the most under-observed. In order to better evaluate performance in the dynamic regions of the scene, we compute a set of dynamic masks for each scene using Track Anything [53]. We compute the intersection between the co-visibility masks and the dynamic regions of the scene and presen… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative evaluation of our method against benchmark methods on the DyCheck test set. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Qualitative evaluation of our ablation study with settings corresponding to Tab. 4. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Qualitative evaluation of our method against benchmark methods on the DyCheck qualitative [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 57 canonical work pages

  1. [1]

    M.-Q. V . Bui, J. Park, J. Oh, and M. Kim. DyBluRF: Dynamic Deblurring Neural Radiance Fields for Blurry Monocular Video. arXiv preprint arXiv:2312.13528, 2023

    arXiv 2023

  2. [2]

    Cao and J

    A. Cao and J. CV . HexPlane: A Fast Representation for Dynamic Scenes. In Computer Vision and Pattern Recognition Conference (CVPR), 2023

    work page 2023

  3. [3]

    Fridovich-Keil, G

    S. Fridovich-Keil, G. Meanti, F. R. Warburg, B. Recht, and A. Kanazawa. K-Planes: Explicit Radiance Fields in Space, Time, and Appearance. In Computer Vision and Pattern Recognition Conference (CVPR), 2023

    work page 2023

  4. [4]

    H. Gao, R. Li, S. Tulsiani, B. Russell, and A. Kanazawa. Monocular Dynamic View Synthesis: A Reality Check. In Conference on Neural Information Processing Systems , 2022

    work page 2022

  5. [5]

    Guizilini, M

    V . Guizilini, M. Z. Irshad, D. Chen, G. Shakhnarovich, and R. Ambrus. Zero-Shot Novel View and Depth Synthesis with Multi-View Geometric Diffusion. In Computer Vision and Pattern Recognition Conference (CVPR), 2025

    work page 2025

  6. [6]

    Huang, Y .-T

    Y .-H. Huang, Y .-T. Sun, Z. Yang, X. Lyu, Y .-P. Cao, and X. Qi. SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes. In Computer Vision and Pattern Recognition Conference (CVPR), 2023

    work page 2023

  7. [7]

    Kappel, F

    M. Kappel, F. Hahlbohm, T. Scholz, S. Castillo, C. Theobalt, M. Eisemann, V . Golyanik, and M. Magnor. D-NPC: Dynamic neural point clouds for non-rigid view synthesis from monocular video. Proceedings of the Eurographics Conference (EG), 44, 2025

    work page 2025

  8. [8]

    Karras, M

    T. Karras, M. Aittala, T. Aila, and S. Laine. Elucidating the design space of diffusion-based generative models. In Conference on Neural Information Processing Systems , 2022

    work page 2022

Show all 64 references
  1. [9]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4), July 2023

    2023

  2. [10]

    J. Lee, C. Won, H. Jung, I. Bae, and H.-G. Jeon. Fully Explicit Dynamic Guassian Splatting. InProceedings of the Neural Information Processing Systems , 2024

    2024

  3. [11]

    J. Lei, Y . Weng, A. Harley, L. Guibas, and K. Daniilidis. MoSca: Dynamic gaussian fusion from casual videos via 4D motion scaffolds. Computer Vision and Pattern Recognition Conference (CVPR), 2025

    2025

  4. [12]

    H. Li, H. Shi, W. Zhang, W. Wu, Y . Liao, L. Wang, L.-h. Lee, and P. Y . Zhou. Dreamscene: 3d gaussian- based text-to-3d scene generation via formation pattern sampling. In European Conference on Computer Vision (ECCV), 2024

    2024

  5. [13]

    L. Li, Z. Shen, Z. Wang, L. Shen, and P. Tan. Streaming Radiance Fields for 3D Video Synthesis. In Conference on Neural Information Processing Systems , 2022

    2022

  6. [14]

    Z. Li, S. Niklaus, N. Snavely, and O. Wang. Neural Scene Flow Fields for Space-Time View Synthesis of Dynamic Scenes. In Computer Vision and Pattern Recognition Conference (CVPR), 2021

    2021

  7. [15]

    Z. Li, Q. Wang, F. Cole, R. Tucker, and N. Snavely. DynIBaR: Neural Dynamic Image-Based Rendering. In Computer Vision and Pattern Recognition Conference (CVPR), 2023

    2023

  8. [16]

    Z. Li, Z. Chen, Z. Li, and Y . Xu. Spacetime Gaussian Feature Splatting for Real-Time Dynamic View Synthesis. In Computer Vision and Pattern Recognition Conference (CVPR), 2024

    2024

  9. [17]

    Liang, J

    H. Liang, J. Ren, A. Mirzaei, A. Torralba, Z. Liu, I. Gilitschenski, S. Fidler, C. Oztireli, H. Ling, Z. Gojcic, and J. Huang. Feed-Forward Bullet-Time Reconstruction of Dynamic Scenes from Monocular Videos. arXiv preprint arXiv:2412.03526, 2024

    2024

  10. [18]

    Liang, X

    Y . Liang, X. Yang, J. Lin, H. Li, X. Xu, and Y . Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In Computer Vision and Pattern Recognition Conference (CVPR), 2024

    2024

  11. [19]

    C. Lin, P. Pan, B. Yang, Z. Li, and Y . Mu. DiffSplat: Repurposing Image Diffusion Models for Scalable 3D Gaussian Splat Generation. In International Conference on Learning Representations (ICLR) , 2025

    2025

  12. [20]

    Y . Lin, Z. Dai, S. Zhu, and Y . Yao. Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle. In Computer Vision and Pattern Recognition Conference (CVPR), 2024. 10

    2024

  13. [21]

    M. Liu, R. Shi, L. Chen, Z. Zhang, C. Xu, X. Wei, H. Chen, C. Zeng, J. Gu, and H. Su. One-2-3-45++: Fast Single Image to 3D Objects with Consistent Multi-View Generation and 3D Diffusion. In Computer Vision and Pattern Recognition Conference (CVPR), 2024

    2024

  14. [22]

    M. Liu, C. Zeng, X. Wei, R. Shi, L. Chen, C. Xu, M. Zhang, Z. Wang, X. Zhang, I. Liu, H. Wu, and H. Su. MeshFormer: High-Quality Mesh Generation with 3D-Guided Reconstruction Model. In Conference on Neural Information Processing Systems, 2024

    2024

  15. [23]

    Q. Liu, Y . Liu, J. Wang, X. Lyu, P. Wang, W. Wang, and J. Hou. MoDGS: Dynamic Gaussian Splatting from Casually-captured Monocular Videos with Depth Priors. In International Conference on Learning Representations (ICLR), 2025

    2025

  16. [24]

    Y .-L. Liu, C. Gao, A. Meuleman, H.-Y . Tseng, A. Saraf, C. Kim, Y .-Y . Chuang, J. Kopf, and J.-B. Huang. Robust dynamic radiance fields. In Computer Vision and Pattern Recognition Conference (CVPR), 2023

    2023

  17. [25]

    Luiten, G

    J. Luiten, G. Kopanas, B. Leibe, and D. Ramanan. Dynamic 3D Gaussians: Tracking by Persistent Dynamic View Synthesis. In International Conference on 3D Vision (3DV), 2024

    2024

  18. [26]

    X. Miao, Y . Bai, H. Duan, F. Wan, Y . Huang, Y . Long, and Y . Zheng. CTNeRF: Cross-time Transformer for dynamic neural radiance field from monocular video. Pattern Recognition, 156:110729, 2024

    2024

  19. [27]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020

    2020

  20. [28]

    Park, M.-Q

    J. Park, M.-Q. V . Bui, J. L. G. Bello, J. Moon, J. Oh, and M. Kim. SplineGS: Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video. In Computer Vision and Pattern Recognition Conference (CVPR), 2025

    2025

  21. [29]

    K. Park, U. Sinha, J. T. Barron, S. Bouaziz, D. B. Goldman, S. M. Seitz, and R. Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. International Conference on Computer Vision (ICCV) , 2021

    2021

  22. [30]

    K. Park, U. Sinha, P. Hedman, J. T. Barron, S. Bouaziz, D. B. Goldman, R. Martin-Brualla, and S. M. Seitz. HyperNeRF: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics, 40(6), 2021

    2021

  23. [31]

    Podell, Z

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, and R. Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. In International Conference on Learning Representations (ICLR), 2024

    2024

  24. [32]

    Poole, A

    B. Poole, A. Jain, J. T. Barron, and B. Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. In International Conference on Learning Representations (ICLR) , 2023

    2023

  25. [33]

    Pumarola, E

    A. Pumarola, E. Corona, G. Pons-Moll, and F. Moreno-Noguer. D-NeRF: Neural Radiance Fields for Dynamic Scenes. In Computer Vision and Pattern Recognition Conference (CVPR), 2021

    2021

  26. [34]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition Conference (CVPR), 2021

    2021

  27. [35]

    N. Ruiz, Y . Li, V . Jampani, Y . Pritch, M. Rubinstein, and K. Aberman. DreamBooth: Fine Tuning Text- to-image Diffusion Models for Subject-Driven Generation. In Computer Vision and Pattern Recognition Conference (CVPR), 2023

    2023

  28. [36]

    R. Shao, Z. Zheng, H. Tu, B. Liu, H. Zhang, and Y . Liu. Tensor4D: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In Computer Vision and Pattern Recognition Conference (CVPR), 2023

    2023

  29. [37]

    R. Shaw, M. Nazarczuk, J. Song, A. Moreau, S. Catley-Chandar, H. Dhamo, and E. Pérez-Pellitero. SWinGS: Sliding Windows for Dynamic 3D Gaussian Splatting. In European Conference on Computer Vision (ECCV), 2024

    2024

  30. [38]

    Shriram, A

    J. Shriram, A. Trevithick, L. Liu, and R. Ramamoorthi. RealmDreamer: Text-Driven 3D Scene Generation with Inpainting and Depth Diffusion. In International Conference on 3D Vision (3DV), 2025

    2025

  31. [39]

    Simonyan and A

    K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations , 2015

    2015

  32. [40]

    Stearns, A

    C. Stearns, A. W. Harley, M. Uy, F. Dubost, F. Tombari, G. Wetzstein, and L. Guibas. Dynamic Gaussian Marbles for Novel View Synthesis of Casual Monocular Videos. In SIGGRAPH Asia, 2024. 11

    2024

  33. [41]

    J. Tang, Z. Chen, X. Chen, T. Wang, G. Zeng, and Z. Liu. LGM: Large Multi-View Gaussian Model for High-Resolution 3D Content Creation. In European Conference on Computer Vision (ECCV), 2024

    2024

  34. [42]

    C. Wang, P. Zhuang, A. Siarohin, J. Cao, G. Qian, H.-Y . Lee, and S. Tulyakov. Diffusion Priors for Dynamic View Synthesis from Monocular Videos. arXiv preprint arXiv:2401.05583, 2024

    2024 arXiv

  35. [43]

    F. Wang, S. Tan, X. Li, Z. Tian, and H. Liu. Mixed Neural V oxels for Fast Multi-view Video Synthesis. In International Conference on Computer Vision (ICCV) , 2023

    2023

  36. [44]

    Q. Wang, V . Ye, H. Gao, W. Zeng, J. Austin, Z. Li, and A. Kanazawa. Shape of Motion: 4D Reconstruction from a Single Video. In arXiv preprint arXiv:2407.13764, 2024

    2024

  37. [45]

    Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4), 2004

    2004

  38. [46]

    Wimmer, M

    T. Wimmer, M. Oechsle, M. Niemeyer, and F. Tombari. Gaussians-to-Life: Text-Driven Animation of 3D Gaussian Splatting Scenes. In International Conference on 3D Vision (3DV), 2025

    2025

  39. [47]

    G. Wu, T. Yi, J. Fang, L. Xie, X. Zhang, W. Wei, W. Liu, Q. Tian, and X. Wang. 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering. In Computer Vision and Pattern Recognition Conference (CVPR), 2024

    2024

  40. [48]

    R. Wu, R. Gao, B. Poole, A. Trevithick, C. Zheng, J. T. Barron, and A. Holynski. CAT4D: Create Anything in 4D with Multi-View Video Diffusion Models. arXiv:2411.18613, 2024

    2024 arXiv

  41. [49]

    R. Wu, B. Mildenhall, P. Henzler, K. Park, R. Gao, D. Watson, P. P. Srinivasan, D. Verbin, J. T. Barron, B. Poole, and A. Holynski. ReconFusion: 3D Reconstruction with Diffusion Priors. In Computer Vision and Pattern Recognition Conference (CVPR), 2024

    2024

  42. [50]

    J. Xu, W. Cheng, Y . Gao, X. Wang, S. Gao, and Y . Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024

    2024 arXiv

  43. [51]

    Y . Xu, H. Tan, F. Luan, S. Bi, P. Wang, J. Li, Z. Shi, K. Sunkavalli, G. Wetzstein, Z. Xu, and K. Zhang. DMV3D: Denoising Multi-View Diffusion using 3D Large Reconstruction Model. In International Conference on Learning Representations (ICLR) , 2024

    2024

  44. [52]

    C. Yang, S. Li, J. Fang, R. Liang, L. Xie, X. Zhang, W. Shen, and Q. Tian. GaussianObject: High-Quality 3D Object Reconstruction from Four Views with Gaussian Splatting. In SIGGRAPH Asia, 2024

    2024

  45. [53]

    J. Yang, M. Gao, Z. Li, S. Gao, F. Wang, and F. Zheng. Track Anything: Segment Anything Meets Videos. arXiv preprint arXiv:2304.11968, 2023

    2023 arXiv

  46. [54]

    S. Yang, W. Yu, J. Zeng, J. Lv, K. Ren, C. Lu, D. Lin, and J. Pang. Novel demonstration generation with gaussian splatting enables robust one-shot manipulation. arXiv preprint arXiv:2504.13175, 2025

    2025 arXiv

  47. [55]

    Z. Yang, X. Gao, W. Zhou, S. Jiao, Y . Zhang, and X. Jin. Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. In Computer Vision and Pattern Recognition Conference (CVPR), 2024

    2024

  48. [56]

    T. Yi, J. Fang, J. Wang, G. Wu, L. Xie, X. Zhang, W. Liu, Q. Tian, and X. Wang. GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models. In Computer Vision and Pattern Recognition Conference (CVPR), 2024

    2024

  49. [57]

    J. S. Yoon, K. Kim, O. Gallo, H. S. Park, and J. Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Computer Vision and Pattern Recognition Conference (CVPR), 2020

    2020

  50. [58]

    H. Yu, C. Wang, P. Zhuang, W. Menapace, A. Siarohin, J. Cao, L. A. Jeni, S. Tulyakov, and H.-Y . Lee. 4Real: Towards Photorealistic 4D Scene Generation via Video Diffusion Models. In Conference on Neural Information Processing Systems, 2024

    2024

  51. [59]

    Y . Zeng, Y . Jiang, S. Zhu, Y . Lu, Y . Lin, H. Zhu, W. Hu, X. Cao, and Y . Yao. STAG4D: Spatial-Temporal Anchored Generative 4D Gaussians. In European Conference on Computer Vision (ECCV), 2024

    2024

  52. [60]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Computer Vision and Pattern Recognition Conference (CVPR), 2018

    2018

  53. [61]

    X. Zhao, A. Colburn, F. Ma, M. Ángel Bautista, J. M. Susskind, and A. G. Schwing. Pseudo-Generalized Dynamic View Synthesis from a Video. In International Conference on Learning Representations (ICLR) , 2024. 12

    2024

  54. [62]

    Zhou, J.-X

    K. Zhou, J.-X. Zhong, S. Shin, K. Lu, Y . Yang, A. Markham, and N. Trigoni. DynPoint: dynamic neural point for view synthesis. In Conference on Neural Information Processing Systems , 2023

    2023

  55. [63]

    R. Zhu, Y . Liang, H. Chang, J. Deng, J. Lu, W. Yang, T. Zhang, and Y . Zhang. MotionGS: Exploring Explicit Motion Guidance for Deformable 3D Gaussian Splatting. In Conference on Neural Information Processing Systems, 2024

    2024

  56. [64]

    Z.-X. Zou, Z. Yu, Y .-C. Guo, Y . Li, D. Liang, Y .-P. Cao, and S.-H. Zhang. Triplane Meets Gaussian Splatting: Fast and Generalizable Single-View 3D Reconstruction with Transformers. In Computer Vision and Pattern Recognition Conference (CVPR), 2024. 13 ViDAR: Video Diffusion...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.