Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Not All Frame Features Are Equal: Video-to-4D Generation via Decoupling Dynamic-Static Features

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

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

Pith's one-line read Decoupling each frame's features into a static projection and a dynamic residual improves video-to-4D generation, especially when static regions dominate.

desk verdict A coherent feature-level dynamic-static decoupling for video-to-4D generation, with consistent ablations but an unvalidated core assumption and thin evaluation. read the letter →

arxiv 2502.08377 v3 pith:RXCLXXO2 submitted 2025-02-12 cs.CV

Liying Yang , Chen Liu , Zhenwei Zhu , Ajian Liu , Hui Ma , Jian Nong , Yanyan Liang This is my paper · ORCID
classification cs.CV
keywords video-to-4Dgenerationdynamic-staticfeaturedecoupling4DGaussiansplattingDINOv2featuresmulti-viewdiffusiontemporal-spatialfusionmotionprediction
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

Video-to-4D generation builds a moving 3D scene from a single video, and this paper claims that the standard recipe of optimizing Gaussians against whole frames loses the motion. When static background occupies most of the frame, whole-frame optimization overfits the static part and renders moving regions blurry. The proposed remedy is to split every frame's visual features, along the temporal axis, into a static component that matches a reference frame and a dynamic component that is the leftover difference, then to fuse the dynamic components across viewpoints before feeding them to the deformation network. The paper reports that this decoupling raises generation quality on the Consistent4D and Objaverse benchmarks and improves real-world dynamic-scene rendering, and it attributes the gains specifically to the dynamic-static decoupling rather than to extra supervision.

What carries the argument

The load-bearing operation is a Gram-Schmidt-style orthogonal split of frame features against a reference. For the $j$-th view at time $i$, the static part is the projection $f^{(i,j)}_{\mathrm{static}} = \left(f^{(i,j)}\cdot r_j / \|r_j\|^2\right) r_j / \|r_j\|$ and the dynamic part is the residual $f^{(i,j)}_{\mathrm{dynamic}} = f^{(i,j)} - f^{(i,j)}_{\mathrm{static}}$. Two references are used: the middle-timestamp frame features, which anchor the scene's semantic content, and the time-averaged frame features, which anchor average motion. The decoupled features are attached to Gaussian points through view projection, and TSSF then fuses them across viewpoints using either global softmax attention or front-view-distance-aware weights; the fused result is combined with HexPlane dynamic features before entering the Deformation MLP. This chain is what converts a 2D feature-space decomposition into a 3D motion signal for 4D generation.

What would settle it

Take a synthetic video of a rigid static object while moving only the light source or the camera; if the DSFD dynamic residual shows high magnitude in regions that did not move, the residual is encoding illumination or viewpoint change rather than motion. A direct version is to compare the spatial heatmap of $f^{(i,j)}_{\mathrm{dynamic}}$ against a ground-truth motion mask from known optical flow and require high overlap in dynamic regions and low overlap in static ones.

Watch

Extended reading notes

Core claim

At the center of the paper is a simple geometric claim about frame features. A frame's feature vector is split by projecting it onto a reference feature $r_j$: the projection $\mathrm{proj}_{r_j}(f^{(i,j)})$ is called the static feature, and the orthogonal residual $f^{(i,j)}_{\mathrm{dynamic}} = f^{(i,j)} - \mathrm{proj}_{r_j}(f^{(i,j)})$ is called the dynamic feature. The paper treats this residual as the carrier of texture, shape, and motion changes between the current frame and the reference, and appends it to the original frame feature so the deformation network sees explicit motion knowledge instead of a frame dominated by static tokens. Together with the temporal-spatial similarity fusion module, which gathers similar dynamic evidence from multiple pseudo-views and weights it adaptively (or by distance to the front view), this decoupling is what the paper identifies as the cause of its reported quality gains.

Load-bearing premise

The load-bearing premise is that, after projecting a frame's features onto the reference features, the leftover orthogonal part is genuinely motion information rather than viewpoint, lighting, or pseudo-multi-view inconsistency; if that premise fails, the DSFD module injects noise instead of motion knowledge.

Editorial extensions

If this is right

  • On the Consistent4D and Objaverse benchmarks, DS4D-GA and DS4D-DA report better CLIP, LPIPS, FVD, and FID-VID scores than Consistent4D, DreamGaussian4D, STAG4D, SC4D, 4Diffusion, and L4GM, with the largest margins on Objaverse's more complex motions.
  • The ablation chain in the paper shows each added piece—frame features, DSFD decoupling, TSSF fusion—improves the metrics, so the reported gains are attributed to the decoupling itself rather than to the point initialization or LPIPS loss alone.
  • Inserting DSFD and TSSF into the 4D-GS baseline on Neu3D real-world scenes improves PSNR, SSIM, and D-SSIM, which the paper reads as evidence that feature-level decoupling transfers to real-world dynamic scenes.
  • The reference-based decoupling is about 14 times faster than decoupling every frame against every other frame, which makes the module practical for optimization pipelines.

Reading between the lines

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

  • Extending beyond the paper, the same residual-decoupling recipe could serve as a cheap motion prior for other video tasks, such as video object segmentation or dynamic scene reconstruction, wherever a reference frame can be chosen.
  • Extending beyond the paper, the method's reliance on pseudo multi-view images means its failure mode is multi-view inconsistency; a natural test is to feed progressively noisier pseudo views and measure whether the dynamic residual tracks motion or the inconsistency.
  • Extending beyond the paper, the DA variant's front-view weighting amounts to a strong prior that the input video is the most reliable view; for videos with large camera motion this prior would need to be re-derived, a setting the paper does not explore.
  • Extending beyond the paper, one could supervise the residual with optical flow or depth from a foundation model to make the static/dynamic split more semantic, an option the paper itself mentions as future work.
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 proposes DS4D, a video-to-4D generation method built on 3D Gaussian splatting. The core idea is to decouple DINOv2 frame features into static and dynamic components by projecting each frame feature onto a reference frame feature along the temporal axis (DSFD), then to fuse dynamic information across pseudo-multi-view images along the spatial axis (TSSF, with GA and DA variants). The decoupled and fused features are used as input to a deformation MLP. The authors report state-of-the-art results on the Consistent4D and Objaverse benchmarks, an ablation study showing incremental gains from each component, and a real-world scene experiment comparing an integrated version of DSFD/TSSF with 4D-GS.

Significance. If the central assumption is valid, the paper offers a lightweight, training-free feature-level decoupling that improves dynamic-region fidelity in optimization-based 4D generation. The ablation table (Table 2) shows consistent improvements when adding DSFD and TSSF, and the robustness experiment (Sec. 12.1) indicates the method degrades gracefully with low-quality multi-view inputs. The modules are clearly described and the code is promised. However, the paper's central claim rests on the orthogonal residual in Eq. (2) being a faithful motion signal, which is asserted rather than validated. The experimental evidence, while positive, is based on single runs over only seven videos or objects per benchmark, and the real-world comparison is internally acknowledged to be unfair. These issues limit the strength of the current SOTA claim.

major comments (3)
  1. [Sec. 3.3, Eq. (2)] The orthogonal residual f_dynamic = f - f_static is asserted to represent dynamic and motion information, but no independent validation is provided. The projection in Eq. (1) is per-token, which implicitly assumes spatial correspondence between the current frame token and the reference frame token; under object motion this correspondence breaks, so the residual can encode misalignment, pseudo-view inconsistencies, lighting changes, or background variation rather than semantic motion. This is load-bearing because Sec. 4.2 attributes all gains to decoupling. Please validate the dynamic-feature interpretation with a concrete test, e.g., comparing the residual to optical flow or ground-truth motion masks on synthetic videos, or replacing the residual with a known motion feature (such as flow-warped features) and showing DSFD performs comparably. The heatmap visualizations (Fig. 8 and Fig. 13) only re-display the constructed features and are not independent evidence.
  2. [Tables 1-2, Sec. 4.2] All quantitative results are reported as single runs on seven Consistent4D videos and seven Objaverse objects, with no error bars, seeds, or significance tests. The claim 'consistently outperform other methods in all metrics' is therefore stronger than the evidence supports; several metric differences (e.g., CLIP and LPIPS in Table 1) are small and could be within run-to-run variation. Please report mean and standard deviation over multiple runs or seeds, or provide per-sample performance to demonstrate statistical robustness. The ablation table 2 shows a consistent trend, which supports the internal mechanism, but the main SOTA claim needs uncertainty quantification.
  3. [Sec. 5 and Supplementary Sec. 10] The real-world experiment is undermined by an internal contradiction. The main text states that for a fair comparison, DSFD and TSSF are inserted into 4D-GS, but the supplementary (Sec. 10, note a) explicitly says 'it is unfair to compare DS4D and 4D-GS due to the different init.' Please clarify whether the reported Ours numbers in Table 3 come from the full DS4D (with LRM initialization) or from a 4D-GS variant with identical initialization, and report the appropriate baseline. As written, the real-world claim is not supported by the admitted confound.
minor comments (5)
  1. [Sec. 4.1, Datasets] The sentence 'In our article, we usefour challenging datasets' contains a missing space; also the list of datasets is incomplete in the main text and would benefit from referencing the supplementary section.
  2. [Supplementary Sec. 10] The phrase 'is unsampled by 2 and 4' should read 'is upsampled by 2 and 4'.
  3. [Eq. (1)] The notation f(i,j) · r_j is ambiguous because f(i,j) is a P x D matrix and r_j is also a matrix; please clarify that the projection is applied per token.
  4. [Sec. 4.2] The sentence 'such significant improvements are attributed to our method of decoupling dynamic-static features' overstates causality; the ablation table also includes changes like LPIPS loss and point initialization, so the attribution should be phrased conditionally.
  5. [Supplementary Sec. 12.1] The robustness experiment compares only against STAG4D; adding another baseline would strengthen the claim that the method is robust to low-quality inputs.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DSFD/TSSF modules are constructional heuristics evaluated on external benchmarks, and the paper's quantitative claims do not reduce to their inputs by definition.

full rationale

The paper's central claim is that decoupling dynamic-static features improves video-to-4D generation. The DSFD module defines static features as the projection of the current token feature onto the reference token feature (Eq. 1) and dynamic features as the orthogonal residual (Eq. 2); this is an explicit modeling assumption rather than a derived result. The subsequent claim that the residual encodes motion is a heuristic premise, not a tautology, because DINOv2 token differences could in principle encode viewpoint or lighting variation. The burden of the claim is carried by the quantitative comparisons in Tables 1 and 2 and the real-world experiments in Table 3, which use external datasets, official baselines, and standard metrics (CLIP, LPIPS, FVD, FID-VID, PSNR, SSIM). These metrics are not defined in terms of the paper's own outputs, so the reported improvements are empirical rather than forced by construction. The heatmap visualizations in Figures 8 and 13 re-display the constructed dynamic features and therefore provide only weak supporting evidence, but this is a limitation of the qualitative validation, not a circular derivation. Self-citations (references [24] and [25]) appear only in the related-work survey and are not load-bearing for the proposed method or its evaluation. The supplementary caveat that the real-world comparison is confounded by different point-cloud initialization methods is a fairness/correctness risk, not a circularity. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from prior work by the same authors, and no external result is replaced by a self-citation chain. The derivation chain is self-contained with respect to its empirical evaluation, so no circular step can be exhibited.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical entities or conserved quantities are proposed. The method combines existing representations: DINOv2 features, Zero123++ pseudo views, LRM point clouds, HexPlane, and 3D Gaussians. The main added assumptions are the semantic interpretation of feature residuals and the heuristic choice of reference frames.

free parameters (4)
  • Loss weights alpha1, alpha2 (Eq. 6) and lambda1-lambda4 (Eq. 7) = not reported
    Tuned by hand; no values or sensitivity analysis appear in the main text or supplementary.
  • Deformation MLP hidden layers and hidden features = 64 hidden layers, 32 hidden features
    Architecture choice stated in Sec. 9; '64 hidden layers' looks like a typo and is not justified.
  • Densification threshold = top 2.5% of points
    Adopted from STAG4D; no sensitivity analysis is provided.
  • Learning rate schedule for TSSF and deformation MLP = 1.6e-4 decaying to 1.6e-6
    Chosen without reported sensitivity analysis.
assumptions (5)
  • ad hoc to paper The orthogonal residual of current frame features after projection onto reference frame features represents dynamic information, and the projected part represents static information.
    Section 3.3, Eqs. (1) and (2). The entire DSFD module rests on this premise, with no independent evidence that residuals encode motion rather than viewpoint or appearance differences.
  • domain assumption Pseudo multi-view images generated by Zero123++ are reliable enough to serve as optimization targets.
    Section 3.2; a footnote defers robustness to the supplementary, and Sec. 12.1 tests only low-quality inputs against one baseline.
  • ad hoc to paper The middle frame feature and the average frame feature adequately represent the semantic and motion content of the whole video.
    Section 3.3 selects these references heuristically, with no criterion or validation for their sufficiency.
  • domain assumption LRM point cloud initialization provides a stable geometric prior.
    Section 3.2; ablation B shows improvement, but this initialization also confounds comparison with baselines that do not use it.
  • domain assumption HexPlane dynamic features regularize 3D Gaussian features and guarantee smoothness.
    Section 3.4, Eq. (5); adopted from prior work [50] without independent verification in this setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Not All Frame Features Are Equal: Video-to-4D Generation via Decoupling Dynamic-Static Features." pith.science (2026). https://pith.science/paper/RXCLXXO2

@misc{pith2026250208377,
  author       = {Pith},
  title        = {Pith review of: Not All Frame Features Are Equal: Video-to-4D Generation via Decoupling Dynamic-Static Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RXCLXXO2}},
  note         = {Machine review of arXiv:2502.08377}
}
read the original abstract

Recently, the generation of dynamic 3D objects from a video has shown impressive results. Existing methods directly optimize Gaussians using whole information in frames. However, when dynamic regions are interwoven with static regions within frames, particularly if the static regions account for a large proportion, existing methods often overlook information in dynamic regions and are prone to overfitting on static regions. This leads to producing results with blurry textures. We consider that decoupling dynamic-static features to enhance dynamic representations can alleviate this issue. Thus, we propose a dynamic-static feature decoupling module (DSFD). Along temporal axes, it regards the regions of current frame features that possess significant differences relative to reference frame features as dynamic features. Conversely, the remaining parts are the static features. Then, we acquire decoupled features driven by dynamic features and current frame features. Moreover, to further enhance the dynamic representation of decoupled features from different viewpoints and ensure accurate motion prediction, we design a temporal-spatial similarity fusion module (TSSF). Along spatial axes, it adaptively selects similar information of dynamic regions. Hinging on the above, we construct a novel approach, DS4D. Experimental results verify our method achieves state-of-the-art (SOTA) results in video-to-4D. In addition, the experiments on a real-world scenario dataset demonstrate its effectiveness on the 4D scene. Our code will be publicly available.

Figures

Figures reproduced from arXiv: 2502.08377 by the authors.

Figure 1
Figure 1. (a) Illustration of the issues caused by different pro￾portions of dynamic and static regions. Previous work [1] gener￾ates the 4D content with obviously blurry textures in the dynamic regions with B-type video input. In contrast, our methods with decoupling dynamic-static features generates high-quality results with clear textures. (b) The visualization of dynamic features in our method. The red region highlights t… view at source ↗
Figure 2
Figure 2. Overview of our proposed DS4D. Given an input video and corresponding multi-view sequences, our DS4D decouples features of the frame at time t based on the reference frame in DSFD module. Next, we acquire point features by retrieving each decoupled feature for Gaussian points via view projection, and we obtain fused Gaussian features by adaptively selecting similar dynamic information from point feature space in TSS… view at source ↗
Figure 3
Figure 3. An overview of the decoupling architecture in DSFD. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: An overview of adaptive fusion in TSSF. views, the effect of merging in GA is affected by that dy￾namic regions are noticeably obscured in specific view￾points. Therefore, we need to reduce the impact of point features under these views for merging while retaining the …
Figure 5
Figure 5. Figure 5: Qualitative comparison on video-to-4D generation. For each method, we render results under two novel views at two timestamps. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 7
Figure 7. Figure 7: Ablation on different experiment settings from Tab.2. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: (a) Visualization on the heatmap of dynamic features in [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison on video-to-4D generation based [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 9
Figure 9. Figure 9: Visualization of flame steak compared with 4D-GS. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Example on low-quality images generated by Image [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Qualitative comparison on 4D generation results with low-quality inputs. For each method, we render results under two novel [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Comparison regarding whether using point clouds gen [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Visualization on the heatmap of dynamic features in [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Visualization on the score map of point features in [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]
Figure 15
Figure 15. Figure 15: Qualitative comparison between model using TSSF [PITH_FULL_IMAGE:figures/full_fig_p015_15.png]
Figure 16
Figure 16. Figure 16: More Results for 4D Generation using DS4D-GA and DS4D-DA. [PITH_FULL_IMAGE:figures/full_fig_p016_16.png]
Figure 17
Figure 17. Figure 17: Qualitative comparison on video-to-4D generation. For each method, we render results under two novel views at two timestamps. [PITH_FULL_IMAGE:figures/full_fig_p017_17.png]
Figure 18
Figure 18. Figure 18: Visualization of real-world 4D scene generation compared with 4D-GS. [PITH_FULL_IMAGE:figures/full_fig_p017_18.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Hallo4D: Multi-Modal Hallucination Mitigation for Consistent Spatio-Temporal Generation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Hallo4D uses vision-language models to detect and correct spatial and temporal mistakes in AI-generated 3D and 4D content, improving consistency without retraining the base generators.

Reference graph

Works this paper leans on

59 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [1]

    Stag4d: Spatial-temporal anchored generative 4d gaussians

    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

    arXiv 2024

  2. [2]

    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 Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 5521–5531, 2022

    2022

  3. [3]

    Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering

    Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632– 16642, 2023

    work page 2023

  4. [4]

    Neural scene flow fields for space-time view synthesis of dy- namic scenes

    Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dy- namic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 6498– 6508, 2021

    work page 2021

  5. [5]

    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. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021

    work page 2021

  6. [6]

    Hypernerf: A higher- dimensional representation for topologically varying neural radiance fields

    Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin- Brualla, and Steven M Seitz. Hypernerf: A higher- dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021

    arXiv 2021

  7. [7]

    Monocular dynamic view synthesis: A reality check

    Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems, 35:33768–33780, 2022

    2022

  8. [8]

    Dreammesh4d: Video-to-4d generation with sparse-controlled gaussian- mesh hybrid representation

    Zhiqi Li, Yiming Chen, and Peidong Liu. Dreammesh4d: Video-to-4d generation with sparse-controlled gaussian- mesh hybrid representation. Advances in Neural Information Processing Systems, 37:21377–21400, 2024

    work page 2024

Show all 59 references
  1. [9]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310–20320, 2024

    2024

  2. [10]

    Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency

    Yiming Xie, Chun-Han Yao, Vikram V oleti, Huaizu Jiang, and Varun Jampani. Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470, 2024

    2024 arXiv

  3. [11]

    4diffusion: Multi-view video diffusion model for 4d generation

    Haiyu Zhang, Xinyuan Chen, Yaohui Wang, Xihui Liu, Yunhong Wang, and Yu Qiao. 4diffusion: Multi-view video diffusion model for 4d generation. arXiv preprint arXiv:2405.20674, 2024

    2024 arXiv

  4. [12]

    L4gm: Large 4d gaus- sian reconstruction model

    Jiawei Ren, Kevin Xie, Ashkan Mirzaei, Hanxue Liang, Xi- aohui Zeng, Karsten Kreis, Ziwei Liu, Antonio Torralba, Sanja Fidler, Seung Wook Kim, et al. L4gm: Large 4d gaus- sian reconstruction model. arXiv preprint arXiv:2406.10324, 2024

    2024 arXiv

  5. [13]

    Diffusion4d: Fast spatial-temporal consis- tent 4d generation via video diffusion models.arXiv preprint arXiv:2405.16645, 2024

    Hanwen Liang, Yuyang Yin, Dejia Xu, Hanxue Liang, Zhangyang Wang, Konstantinos N Plataniotis, Yao Zhao, and Yunchao Wei. Diffusion4d: Fast spatial-temporal consis- tent 4d generation via video diffusion models.arXiv preprint arXiv:2405.16645, 2024

    2024 arXiv

  6. [14]

    Text-to-4d dy- namic scene generation

    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 , 2023

    2023 arXiv

  7. [15]

    4dgen: Grounded 4d content gen- eration with spatial-temporal consistency

    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

    2023 arXiv

  8. [16]

    Sc4d: Sparse-controlled video- to-4d generation and motion transfer

    Zijie Wu, Chaohui Yu, Yanqin Jiang, Chenjie Cao, Fan Wang, and Xiang Bai. Sc4d: Sparse-controlled video- to-4d generation and motion transfer. arXiv preprint arXiv:2404.03736, 2024

    2024 arXiv

  9. [17]

    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. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023

    2023

  10. [18]

    Zero123++: a single image to consistent multi-view dif- fusion base model

    Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view dif- fusion base model. arXiv preprint arXiv:2310.15110, 2023

    arXiv

  11. [19]

    Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models

    Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191 , 2024

    2024 arXiv

  12. [20]

    Consistent4d: Consistent 360° dynamic object gen- eration from monocular video

    Yanqin Jiang, Li Zhang, Jin Gao, Weiming Hu, and Yao Yao. Consistent4d: Consistent 360° dynamic object gen- eration from monocular video. In The Twelfth International Conference on Learning Representations, 2024

    2024

  13. [21]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

    2023

  14. [22]

    3d-r2n2: A unified approach for single and multi-view 3d object reconstruction

    Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision , pages 628–644. Springer, 2016

    2016

  15. [23]

    Pix2vox++: Multi-scale context- aware 3d object reconstruction from single and multi- ple images

    Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context- aware 3d object reconstruction from single and multi- ple images. International Journal of Computer Vision , 128(12):2919–2935, 2020

    2020

  16. [24]

    Umiformer: Mining the correlations between similar tokens for multi-view 3d reconstruction

    Zhenwei Zhu, Liying Yang, Ning Li, Chaohao Jiang, and Yanyan Liang. Umiformer: Mining the correlations between similar tokens for multi-view 3d reconstruction. In Proceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 18226–18235, 2023

    2023

  17. [25]

    Long-range grouping transformer for multi- view 3d reconstruction

    Liying Yang, Zhenwei Zhu, Xuxin Lin, Jian Nong, and Yanyan Liang. Long-range grouping transformer for multi- view 3d reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18257– 18267, 2023

    2023

  18. [26]

    Pixel2mesh: Generating 3d mesh models from single rgb images

    Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the Euro- pean conference on computer vision (ECCV) , pages 52–67, 2018

    2018

  19. [27]

    Differentiable volumetric rendering: Learn- ing implicit 3d representations without 3d supervision

    Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learn- ing implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 3504–3515, 2020

    2020

  20. [28]

    Pixel2mesh++: 3d mesh generation and refinement from multi-view images

    Chao Wen, Yinda Zhang, Chenjie Cao, Zhuwen Li, Xi- angyang Xue, and Yanwei Fu. Pixel2mesh++: 3d mesh generation and refinement from multi-view images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2166–2180, 2022

    2022

  21. [29]

    Multires- olution tree networks for 3d point cloud processing

    Matheus Gadelha, Rui Wang, and Subhransu Maji. Multires- olution tree networks for 3d point cloud processing. In Pro- ceedings of the European Conference on Computer Vision (ECCV), pages 103–118, 2018

    2018

  22. [30]

    Learning representations and generative models for 3d point clouds

    Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40–49. PMLR, 2018

    2018

  23. [31]

    Diffusion probabilistic models for 3d point cloud generation

    Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 2837–2845, 2021

    2021

  24. [32]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    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. Communications of the ACM, 65(1):99–106, 2021

    2021

  25. [33]

    Instant neural graphics primitives with a mul- tiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1–15, 2022

    2022

  26. [34]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1, 2023

    2023

  27. [35]

    Mip-splatting: Alias-free 3d gaussian splat- ting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splat- ting. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 19447–19456, 2024

    2024

  28. [36]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information...

    2022

  29. [37]

    Barron, and Ben Milden- hall

    Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representa- tions, 2023

    2023

  30. [38]

    Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation. InThe Twelfth International Con- ference on Learning Representations, 2024

    2024

  31. [39]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865–5874, 2021

    2021

  32. [40]

    Non- rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video

    Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollh¨ofer, Christoph Lassner, and Christian Theobalt. Non- rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Con...

    2021

  33. [41]

    Dˆ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video

    Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, For- rester Cole, and Cengiz Oztireli. Dˆ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in neural information processing systems , 35:32653–32666, 2022

    2022

  34. [42]

    Fast dynamic radiance fields with time-aware neural vox- els

    Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xi- aopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural vox- els. In SIGGRAPH Asia 2022 Conference Papers, pages 1–9, 2022

    2022

  35. [43]

    Spacetime gaus- sian feature splatting for real-time dynamic view synthesis

    Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaus- sian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508–8520, 2024

    2024

  36. [44]

    Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis

    Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen- Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv preprint arXiv:2312.11458 , 2023

    2023 arXiv

  37. [45]

    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. In 2024 International Con- ference on 3D Vision (3DV), pages 800–809. IEEE, 2024

    2024

  38. [46]

    Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting

    Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023

    2023 arXiv

  39. [47]

    Dreamgaussian4d: Genera- tive 4d gaussian splatting

    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, 2023

    2023 arXiv

  40. [48]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  41. [49]

    Pc2: Projection-conditioned point cloud diffu- sion for single-image 3d reconstruction

    Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffu- sion for single-image 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12923–12932, 2023

    2023

  42. [50]

    Hexplane: A fast representa- tion for dynamic scenes

    Ang Cao and Justin Johnson. Hexplane: A fast representa- tion for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 130–141, 2023

    2023

  43. [51]

    Lrm: Large reconstruction model for single image to 3d

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023

    2023 arXiv

  44. [52]

    Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers

    Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern...

    2024

  45. [53]

    Depthsplat: Connecting gaussian splatting and depth

    Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024

    2024 arXiv

  46. [54]

    Depth any- thing v2

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

    2024 arXiv

  47. [55]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018

    2018

  48. [56]

    Dreamgaussian: Generative gaussian splatting for effi- cient 3d content creation

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for effi- cient 3d content creation. arXiv preprint arXiv:2309.16653, 2023

    2023 arXiv

  49. [57]

    Very deep convolutional networks for large-scale image recognition

    Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014

    2014 arXiv

  50. [58]

    Stable video diffusion: Scaling latent video diffusion models to large datasets

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023

    2023 arXiv

  51. [59]

    Imagedream: Image-prompt multi-view diffusion for 3d generation

    Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023

    2023 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.