Pith. sign in

REVIEW 4 major objections 5 minor 60 references

4DVD: Cascaded Dense-view Video Diffusion Model for High-quality 4D Content Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read 4DVD claims that splitting 4D generation into a 16-view layout stage and a structure-conditioned refinement stage produces multi-view videos consistent enough to reconstruct accurate 4D Gaussians from a single monocular clip.

desk verdict A credible two-stage method whose '16-view' headline outruns its evidence: all direct metrics use only 4 views, and the fine stage is trained on 4 but used on 16. read the letter →

arxiv 2508.04467 v1 pith:4XXE2C3I submitted 2025-08-06 cs.CV

Shuzhou Yang , Xiaodong Cun , Xiaoyu Li , Yaowei Li , Jian Zhang This is my paper
classification cs.CV
keywords 4Dgenerationmulti-viewvideodiffusioncascadedmonocularGaussiandenseviewsD-Objaversestructure-aware
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

4DVD claims that generating high-quality 4D content from a single monocular video becomes tractable when the problem is split into two learned stages rather than modeled end-to-end. In the first stage, a video diffusion model produces a low-resolution 16-view video that locks down consistent geometry and motion; in the second, a frozen base model conditioned on those coarse layouts and on a Monocular Appearance Propagation module upgrades them to high-resolution, appearance-faithful dense-view videos. The paper argues that because spatial detail is redundant across views and frames, resolution can be traded for viewpoint density, and that this trade-off is what makes explicit 4D reconstruction, such as 4D Gaussian optimization, accurate. If the claim holds, the practical payoff is that everyday monocular videos could yield full 360-degree, time-consistent 4D assets at lower cost and with better consistency than existing sparse-view generation pipelines.

What carries the argument

Two-stage cascaded latent video diffusion. Stage 1 finetunes SV4D with LoRA on 16-view 256x256 videos to output dense-view layout priors, sacrificing resolution to fit dense viewpoints in VRAM. Stage 2 keeps SV4D frozen and adds a trainable structure condition branch (with zero-conv injections) plus the Monocular Appearance Propagation (MAP) module, a swapped-Q/KV cross-attention: the sharp monocular reference video supplies the query matrix while the coarse multi-view layouts supply keys and values, propagating appearance from one view to novel views. The workhorse insight is that spatial information is over-redundant across views and time, so the model can learn better 3D structure from mo

What would settle it

Run 4DVD on real, hand-held monocular videos of dynamic objects (e.g., people walking, waving flags) with multi-view ground truth captured by a synchronized camera rig, and measure FVD-V and LPIPS against SV4D and 4Diffusion on held-out real views; if cross-view consistency or reconstruction quality degrades sharply relative to the synthetic D-Objaverse numbers, the synthetic-generalization premise fails. A cheaper probe: reconstruct a 4D Gaussian from 4DVD's 16 outputs on a real clip and check for free-view floaters or ghosting that the in-dataset results do not exhibit.

Watch

Extended reading notes

Core claim

The paper establishes that decoupling 4D generation into coarse dense-view layout prediction followed by structure-aware spatio-temporal generation enables, for the first time, 16-view multi-view video generation from a monocular video. Compared with prior methods that directly model the T-by-V video grid with limited viewpoints (4 views in 4Diffusion, 8 in SV4D), the two-stage cascade produces videos with superior cross-view and temporal consistency, and the dense views provide sufficient supervision for accurate optimization of explicit 4D representations such as 4D Gaussian. The authors claim this is why 4DVD outperforms all baselines on both multi-view video quality (LPIPS, CLIP-S, FVD v

Load-bearing premise

The load-bearing premise is that a model trained and evaluated entirely on clean 16-view synthetic renders from the filtered D-Objaverse dataset will transfer to the real-world monocular videos promised in the introduction; no real-video test appears anywhere in the paper, so out-of-distribution footage could invalidate the practical claim.

Editorial extensions

If this is right

  • 16-view generation enables accurate optimization of explicit 4D representations like 4D Gaussian, which sparse-view outputs are too inconsistent to support; reconstructed 4D assets from 4DVD beat STAG4D, 4Diffusion, L4GM, and SV4D on all reported metrics.
  • The cascaded design is faster than the one-stage SV4D baseline (381 s vs 565 s on an A6000), because both stages need only 10 sampling steps when the coarse layout already anchors the structure.
  • The ablations show that neither ControlNet nor T2I-Adapter condition branches reproduce the quality of the proposed layer-by-layer structure branch, and that the swapped Q/KV design of MAP is necessary: conventional cross-attention introduces artifacts and color offset.
  • The newly collected D-Objaverse dataset, 17k high-quality 16-view dynamic-object videos filtered from Objaverse, makes dense-view training practical and will be released with the paper.
  • The same two-stage decoupling is expected to transfer to diffusion-transformer backbones, which the authors name as the path to scaling to more complex geometry and motion.

Reading between the lines

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

  • The decoupling principle should transfer beyond the pixel-space U-Net used here: diffusion transformers have even heavier memory demands for cross-view/temporal attention, so the coarse-layout-then-refine strategy may become the standard way to reach dense viewpoints at scale.
  • Because all training and evaluation happens on synthetic rendered objects from D-Objaverse, the central practical promise—working from 'widely available monocular videos'—remains untested; a real-video benchmark would separate the method's contribution from the dataset's cleanliness.
  • The MAP swapped-Q/KV attention could generalize to other tasks where a high-quality reference signal must be aligned with a misaligned multi-view condition, such as scene-level view extrapolation or reference-based video editing, not just object-centric 4D.
  • The paper's own limitation note—weak performance on complex geometry and motion—suggests the next bottleneck will be training-data coverage rather than the cascade itself, so a larger or more diverse dense-view dataset could extend the same architecture to hard cases.
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 4DVD, a cascaded video diffusion model for generating dense-view (16-view) multi-view videos from a monocular video, intended to support 4D content reconstruction. The method decouples the problem into two stages: (1) a coarse stage that fine-tunes SV4D with LoRA on low-resolution 16-view videos to predict consistent multi-view layouts, and (2) a fine stage that freezes a base SV4D and trains a structure-condition branch with a proposed Monocular Appearance Propagation (MAP) module to generate high-resolution multi-view videos conditioned on the coarse layouts. The authors collect and filter a dynamic 3D object dataset, D-Objaverse, from Objaverse, and report quantitative and qualitative comparisons against STAG4D, 4Diffusion, L4GM, and SV4D on multi-view video synthesis and 4D Gaussian reconstruction.

Significance. If the dense-view consistency claim holds, 4DVD would be a useful step toward practical 4D content generation from monocular video, since it is the first pipeline to explicitly generate 16-view videos within a video-diffusion framework and use them for explicit 4D reconstruction. The cascaded coarse-to-fine design is well motivated, and the MAP attention mechanism is an interesting architectural contribution. The release of a filtered high-quality dynamic dataset is also valuable. The runtime comparison shows that the cascade is competitive with (indeed faster than) the single-stage SV4D baseline. However, the central claim—that the 16-view outputs are sufficiently consistent to improve 4D reconstruction—is not directly measured, and several train/inference mismatches are left unexamined.

major comments (4)
  1. [Sec. 5.2, Tab. 1] The paper's headline claim is that 16-view generation yields more consistent multi-view videos and enables high-quality 4D reconstruction. Yet all multi-view video metrics in Tab. 1 are computed only on the T×4 orthogonal-view grid ('we only use videos from the four orthogonal viewpoints for comparison'). The 12 non-orthogonal views—the actual novel output of 4DVD—are excluded from every video metric. The 4D-asset metrics presumably use all 16 views, but the manuscript does not state how many views are used for the 4D Gaussian optimization or how the extra views are weighted. Please report direct metrics on the full 16-view grid (e.g., LPIPS/CLIP-S/FVD on non-orthogonal views, or per-view consistency errors), and clarify the reconstruction input. Without this, the reported superiority could be driven entirely by the four best views and the dense-view advantage remains unverified.
  2. [Sec. 3.4, Sec. 5.1, Eq. (5)] There is an unexamined train/inference mismatch in the fine stage. The text states that during training, 'we randomly select 4 views of the 576×576 resolution at each training iteration,' while at inference the model predicts the full 16-view grid. The assertion that 4-view training suffices for 16-view inference is plausible but not tested. More importantly, the training conditions in Eq. (5) are the ground-truth low-resolution renders, whereas during inference the conditions are the stage-1 predicted layouts. No condition dropout, noise augmentation, or other robustness mechanism is described to bridge this gap between clean training conditions and noisy predicted conditions. Please provide an ablation that compares (a) train with more views or (b) inference with 4-view vs 16-view conditions, and report the effect of conditioning on stage-1 outputs versus ground-truth layouts. This is
  3. [Sec. 4, Sec. 5.1, Intro] All training and evaluation is performed on D-Objaverse, a filtered subset of synthetic Objaverse renders. The introduction explicitly promises generation from 'widely available monocular videos' for AR/VR creation, but no experiment tests real-world videos. Synthetic renders have clean backgrounds, known camera trajectories, and no motion blur or occlusions. Since the model is trained only on such data, the claim of practical applicability is unsupported. At minimum, include a qualitative or quantitative test on real monocular videos (e.g., DAVIS or object-centric YouTube videos), or explicitly scope the paper's claims to synthetic inputs. This does not invalidate the in-dataset results, but it is essential for the stated motivation.
  4. [Tab. 1, Tab. 3] The quantitative comparisons report single point estimates on 50 test objects with no error bars or significance tests. The improvement over SV4D in LPIPS (0.136 vs 0.133) and CLIP-S (0.913 vs 0.927) is small, and without confidence intervals it is impossible to know whether the reported gains are robust. The same issue affects the ablation table. Please report standard errors, bootstrap confidence intervals, or paired significance tests for the main comparisons. This is important because the paper's central claim of consistent dense-view superiority rests on these marginal differences.
minor comments (5)
  1. [Eq. (3)] The dimension notation in Eq. (3) is confusing: 'W H V' appears to mean W×H×V, but W and H are not defined in this equation and the text later says 'W H and Nc mean the spatial resolution and feature channel number.' Please clarify the exact shapes of Q, K, V and the folding/unfolding of the view dimension.
  2. [Sec. 4, Sec. 5.2, Fig. 7] Several typos: 'dnamic' (Sec. 4), 'quantitaive' (Sec. 5.2), 'Similiar' (Fig. 7 caption), 'T able' in table captions. The typo 'D-objaverse' vs 'D-Objaverse' is inconsistent.
  3. [Fig. 3] The figure label 'SV4D Stage 1GT' is unclear; it should be three separate labels (SV4D, Stage 1, GT) or a legend. The residual images are not described in the caption; please specify how residuals are computed.
  4. [Sec. 3.3] The first paragraph claims that SV4D 'fails to predict accurate results when the number of viewpoints is greater than the number of views in the training set.' This is supported only with a single visual example (Fig. 3). A quantitative residual comparison or a small experiment would strengthen the motivation for the dense-view coarse stage.
  5. [Sec. 5.5, Tab. 3] The 'MAP w/ feature' ablation is described as replacing MAP with 'common cross-attention' where the current model features are used as Q. Please specify whether this variant still uses the reference video as K/V, and why the dimensions are compatible. A one-sentence architectural detail would make the ablation reproducible.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: 4DVD's two-stage generation is benchmarked against external baselines; the only self-citations are minor and non-load-bearing.

full rationale

No circular derivation chain was found. 4DVD is a cascade of two separately trained stages: Stage 1 finetunes SV4D with LoRA on 16-view low-resolution videos to predict coarse layouts (Eq. 4), and Stage 2 trains a structure-condition branch with the proposed MAP module against high-resolution videos conditioned on low-resolution layouts (Eq. 5), with the base model frozen. The final 16-view output is therefore a learned composition of the monocular input and the Stage-1 layouts, not a restatement of either training objective. The 4D-Gaussian results are produced by an external reconstruction method (DreamGaussian4D) from the generated videos, so the downstream evaluation is not identical to the generation loss. The central claims are tested against external baselines (STAG4D, 4Diffusion, L4GM, SV4D) and ablated against ControlNet, T2I-adapter, and a conventional cross-attention variant, giving the method independent evidential content. The only self-citations are [22] and [58], used in related-work enumeration and as an example of a conventional cross-attention baseline; neither is load-bearing for the paper's claims. Concerns that multi-view metrics are computed only on the four orthogonal views and that Stage 2 is trained with V=4 while inferring at V=16 are evaluation-validity or generalization issues, not cases where a prediction reduces to its input by construction. Thus the appropriate circularity score is low; the minor self-citations are non-load-bearing, so the score is 2 rather than 0.

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

The central claim rests on hand-chosen operating points (16 views, 256x256 coarse resolution, 4 training views, T=5) and on domain assumptions about the reliability of external priors (SV3D reference views, frozen SV4D base, VAE behavior at low resolution) and about the representativeness of the filtered D-Objaverse dataset. No new physical or conceptual entities are introduced.

free parameters (4)
  • number of dense views V = 16
    Core to the claim of dense-view generation; chosen by hand as 'dense enough' coverage of a 4D object (Sec 3.3), not derived.
  • coarse-stage resolution = 256x256
    Resolution sacrificed to enable 16-view LoRA training; justified by VAE limitations and experiments (Sec 3.3), but is a hand-chosen operating point.
  • fine-stage training views = 4
    Randomly select 4 of 16 views per training iteration to fit VRAM; inference uses all 16 (Sec 3.4, 5.1), a train/inference mismatch.
  • sequence length T = 5
    Training uses 5-frame clips; 21-frame videos are stitched via SV4D anchor sampling (Sec 5.1).
assumptions (4)
  • domain assumption SV3D provides sufficiently accurate reference views for the first frame
    Stage 1 takes 16 reference views synthesized by SV3D [51]; if these are wrong, the layout priors encode wrong geometry (Sec 3.3).
  • domain assumption SV4D can serve as a frozen base model for 16-view generation after branch conditioning
    Stage 2 freezes SV4D and only trains a control branch; the base model's attention mechanisms must generalize to 16 views at inference despite being trained on 4 (Sec 3.4).
  • domain assumption VAE encoder-decoder preserves 3D layout structure at 256x256
    Stage 1 relies on low-res latents retaining enough structure to guide dense-view generation (Sec 3.3).
  • domain assumption The LPIPS/optical-flow/CogVLM filtering yields a high-quality, unbiased sample of dynamic objects
    D-Objaverse filtering defines training and test sets; biased filters could inflate evaluation scores (Sec 4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of 4DVD: Cascaded Dense-view Video Diffusion Model for High-quality 4D Content Generation." pith.science (2026). https://pith.science/paper/4XXE2C3I

@misc{pith2026250804467,
  author       = {Pith},
  title        = {Pith review of: 4DVD: Cascaded Dense-view Video Diffusion Model for High-quality 4D Content Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4XXE2C3I}},
  note         = {Machine review of arXiv:2508.04467}
}
read the original abstract

Given the high complexity of directly generating high-dimensional data such as 4D, we present 4DVD, a cascaded video diffusion model that generates 4D content in a decoupled manner. Unlike previous multi-view video methods that directly model 3D space and temporal features simultaneously with stacked cross view/temporal attention modules, 4DVD decouples this into two subtasks: coarse multi-view layout generation and structure-aware conditional generation, and effectively unifies them. Specifically, given a monocular video, 4DVD first predicts the dense view content of its layout with superior cross-view and temporal consistency. Based on the produced layout priors, a structure-aware spatio-temporal generation branch is developed, combining these coarse structural priors with the exquisite appearance content of input monocular video to generate final high-quality dense-view videos. Benefit from this, explicit 4D representation~(such as 4D Gaussian) can be optimized accurately, enabling wider practical application. To train 4DVD, we collect a dynamic 3D object dataset, called D-Objaverse, from the Objaverse benchmark and render 16 videos with 21 frames for each object. Extensive experiments demonstrate our state-of-the-art performance on both novel view synthesis and 4D generation. Our project page is https://4dvd.github.io/

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 37 canonical work pages

  1. [1]

    In: The Twelfth International Conference on Learning Representations (2024)

    Jiang, Y., Zhang, L., Gao, J., Hu, W., Yao, Y.: Consistent4d: Consistent 360 ° dynamic object generation from monocular video. In: The Twelfth International Conference on Learning Representations (2024)

    work page 2024

  2. [2]

    In: Proceedings of the 40th International Conference on Machine Learning (2023)

    Singer, U., Sheynin, S., Polyak, A., Ashual, O., Makarov, I., Kokkinos, F., Goyal, N., Vedaldi, A., Parikh, D., Johnson, J., Taig- man, Y.: Text-to-4d dynamic scene genera- tion. In: Proceedings of the 40th International Conference on Machine Learning (2023)

    work page 2023

  3. [3]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    Bahmani, S., Skorokhodov, I., Rong, V., Wet- zstein, G., Guibas, L., Wonka, P., Tulyakov, S., Park, J.J., Tagliasacchi, A., Lindell, D.B.: 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    work page 2024

  4. [4]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text- to-4d with dynamic 3d gaussians and com- posed diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    work page 2024

  5. [5]

    arXiv preprint arXiv:2312.17142 (2024)

    Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Gen- erative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2024)

    arXiv 2024

  6. [6]

    arXiv preprint arXiv:2407.17470 (2024)

    Xie, Y., Yao, C.-H., Voleti, V., Jiang, H., Jampani, V.: Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470 (2024)

    arXiv 2024

  7. [7]

    arXiv preprint arXiv:2406.10324 (2024) Springer Nature 2022 LATEX template Article Title 15

    Ren, J., Xie, K., Mirzaei, A., Liang, H., Zeng, X., Kreis, K., Liu, Z., Torralba, A., Fidler, S., Kim, S.W., Ling, H.: L4gm: Large 4d gaus- sian reconstruction model. arXiv preprint arXiv:2406.10324 (2024) Springer Nature 2022 LATEX template Article Title 15

    arXiv 2024

  8. [8]

    In: Advances in Neural Information Processing Systems (2024)

    Yu, H., Wang, C., Zhuang, P., Menapace, W., Siarohin, A., Cao, J., Jeni, L.A., Tulyakov, S., Lee, H.-Y.: 4real: Towards photorealis- tic 4d scene generation via video diffusion models. In: Advances in Neural Information Processing Systems (2024)

    work page 2024

Show all 60 references
  1. [9]

    arXiv preprint arXiv:2405.20674 (2024)

    Zhang, H., Chen, X., Wang, Y., Liu, X., Wang, Y., Qiao, Y.: 4diffusion: Multi-view video diffusion model for 4d generation. arXiv preprint arXiv:2405.20674 (2024)

    2024 arXiv

  2. [10]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

    Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Obja- verse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

    2023

  3. [11]

    In: The Eleventh International Confer- ence on Learning Representations (2023)

    Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffu- sion. In: The Eleventh International Confer- ence on Learning Representations (2023)

    2023

  4. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

    Lin, C.-H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.-Y., Lin, T.-Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

    2023

  5. [13]

    In: Advances in Neu- ral Information Processing Systems (2023)

    Wang, Z., Lu, C., Wang, Y., Bao, F., LI, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with varia- tional score distillation. In: Advances in Neu- ral Information Processing Systems (2023)

    2023

  6. [14]

    arXiv preprint arXiv:2310.06744 (2024)

    Yu, W., Yuan, L., Cao, Y.-P., Gao, X., Li, X., Hu, W., Quan, L., Shan, Y., Tian, Y.: Hifi-123: Towards high-fidelity one image to 3d content generation. arXiv preprint arXiv:2310.06744 (2024)

    2024 arXiv

  7. [15]

    arXiv preprint arXiv:2310.08092 (2023)

    Weng, H., Yang, T., Wang, J., Li, Y., Zhang, T., Chen, C.L.P., Zhang, L.: Con- sistent123: Improve consistency for one image to 3d object synthesis. arXiv preprint arXiv:2310.08092 (2023)

    2023 arXiv

  8. [16]

    arXiv preprint arXiv:2310.08529 (2024)

    Yi, T., Fang, J., Wang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffu- sion models. arXiv preprint arXiv:2310.08529 (2024)

    2024 arXiv

  9. [17]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    Chen, Z., Wang, F., Wang, Y., Liu, H.: Text-to-3d using gaussian splatting. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    2024

  10. [18]

    In: The Twelfth International Confer- ence on Learning Representations (2024)

    Sun, J., Zhang, B., Shao, R., Wang, L., Liu, W., Xie, Z., Liu, Y.: Dreamcraft3d: Hierarchi- cal 3d generation with bootstrapped diffusion prior. In: The Twelfth International Confer- ence on Learning Representations (2024)

    2024

  11. [19]

    In: The Twelfth International Conference on Learn- ing Representations (2024)

    Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splat- ting for efficient 3d content creation. In: The Twelfth International Conference on Learn- ing Representations (2024)

    2024

  12. [20]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

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

    2024

  13. [21]

    In: Proceed- ings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (2024)

    Zhou, L., Shih, A., Meng, C., Ermon, S.: Dreampropeller: Supercharge text-to-3d gen- eration with parallel sampling. In: Proceed- ings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (2024)

    2024

  14. [22]

    arXiv preprint arXiv:2405.20669 (2024)

    Yang, S., Wang, Y., Li, H., Meng, J., Wu, Y., Meng, X., Zhang, J.: Hybrid fourier score distillation for efficient one image to 3d object generation. arXiv preprint arXiv:2405.20669 (2024)

    2024 arXiv

  15. [23]

    In: The Twelfth Inter- national Conference on Learning Representa- tions (2024)

    Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: LRM: Large reconstruction model for single image to 3d. In: The Twelfth Inter- national Conference on Learning Representa- tions (2024)

    2024

  16. [24]

    In: The Twelfth Inter- national Conference on Learning Representa- tions (2024)

    Jiang, H., Jiang, Z., Zhao, Y., Huang, Q.: LEAP: Liberate sparse-view 3d modeling Springer Nature 2022 LATEX template 16 Article Title from camera poses. In: The Twelfth Inter- national Conference on Learning Representa- tions (2024)

    2022

  17. [25]

    In: The Twelfth International Conference on Learning Representations (2024)

    Wang, P., Tan, H., Bi, S., Xu, Y., Luan, F., Sunkavalli, K., Wang, W., Xu, Z., Zhang, K.: PF-LRM: Pose-free large reconstruction model for joint pose and shape prediction. In: The Twelfth International Conference on Learning Representations (2024)

    2024

  18. [26]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (2024)

    Zou, Z.-X., Yu, Z., Guo, Y.-C., Li, Y., Liang, D., Cao, Y.-P., Zhang, S.-H.: Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transform- ers. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (2024)

    2024

  19. [27]

    arXiv preprint arXiv:2404.12385 (2024)

    Wei, X., Zhang, K., Bi, S., Tan, H., Luan, F., Deschaintre, V., Sunkavalli, K., Su, H., Xu, Z.: Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385 (2024)

    2024 arXiv

  20. [28]

    arXiv preprint arXiv:2403.02151 (2024)

    Tochilkin, D., Pankratz, D., Liu, Z., Huang, Z., Letts, A., Li, Y., Liang, D., Laforte, C., Jampani, V., Cao, Y.-P.: Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151 (2024)

    2024 arXiv

  21. [29]

    In: European Conference on Computer Vision (2025)

    Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In: European Conference on Computer Vision (2025)

    2025

  22. [30]

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: representing scenes as neural radiance fields for view synthesis. Commun. ACM (2021)

    2021

  23. [31]

    arXiv preprint arXiv:2308.04079 (2023)

    Kerbl, B., Kopanas, G., Leimk¨ uhler, T., Dret- takis, G.: 3d gaussian splatting for real- time radiance field rendering. arXiv preprint arXiv:2308.04079 (2023)

    2023 arXiv

  24. [32]

    In: The Twelfth Interna- tional Conference on Learning Representa- tions (2024)

    Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: MVDream: Multi-view diffusion for 3d generation. In: The Twelfth Interna- tional Conference on Learning Representa- tions (2024)

    2024

  25. [33]

    arXiv preprint arXiv:2312.02201 (2023)

    Wang, P., Shi, Y.: Imagedream: Image- prompt multi-view diffusion for 3d genera- tion. arXiv preprint arXiv:2312.02201 (2023)

    2023 arXiv

  26. [34]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    Long, X., Guo, Y.-C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.-H., Habermann, M., Theobalt, C., Wang, W.: Wonder3d: Sin- gle image to 3d using cross-domain diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    2024

  27. [35]

    arXiv preprint arXiv:2405.11616 (2024)

    Li, P., Liu, Y., Long, X., Zhang, F., Lin, C., Li, M., Qi, X., Zhang, S., Luo, W., Tan, P., Wang, W., Liu, Q., Guo, Y.: Era3d: High-resolution multiview diffusion using efficient row-wise attention. arXiv preprint arXiv:2405.11616 (2024)

    2024 arXiv

  28. [36]

    arXiv preprint arXiv:2311.14603 (2024)

    Zhao, Y., Yan, Z., Xie, E., Hong, L., Li, Z., Lee, G.H.: Animate124: Animating one image to 4d dynamic scene. arXiv preprint arXiv:2311.14603 (2024)

    2024 arXiv

  29. [37]

    arXiv preprint arXiv:2312.17225 (2024)

    Yin, Y., Xu, D., Wang, Z., Zhao, Y., Wei, Y.: 4dgen: Grounded 4d content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225 (2024)

    2024 arXiv

  30. [38]

    In: European Conference on Computer Vision (2025)

    Zeng, Y., Jiang, Y., Zhu, S., Lu, Y., Lin, Y., Zhu, H., Hu, W., Cao, X., Yao, Y.: Stag4d: Spatial-temporal anchored genera- tive 4d gaussians. In: European Conference on Computer Vision (2025)

    2025

  31. [39]

    arXiv preprint arXiv:2401.08742 (2024)

    Pan, Z., Yang, Z., Zhu, X., Zhang, L.: Effi- cient4d: Fast dynamic 3d object generation from a single-view video. arXiv preprint arXiv:2401.08742 (2024)

    2024

  32. [40]

    arXiv preprint arXiv:2404.02148 (2024)

    Yang, Z., Pan, Z., Gu, C., Zhang, L.: Diffusion2: Dynamic 3d content genera- tion via score composition of video and multi-view diffusion models. arXiv preprint arXiv:2404.02148 (2024)

    2024 arXiv

  33. [41]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) Springer Nature 2022 LATEX template Article Title 17

    Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) Springer Nature 2022 LATEX template Article Title 17

    2021

  34. [42]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

    Cao, A., Johnson, J.: Hexplane: A fast repre- sentation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

    2023

  35. [43]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    2024

  36. [44]

    arXiv preprint arXiv:2405.16645 (2024)

    Liang, H., Yin, Y., Xu, D., Liang, H., Wang, Z., Plataniotis, K.N., Zhao, Y., Wei, Y.: Dif- fusion4d: Fast spatial-temporal consistent 4d generation via video diffusion models. arXiv preprint arXiv:2405.16645 (2024)

    2024 arXiv

  37. [45]

    arXiv preprint arXiv:2411.04928 (2024)

    Sun, W., Chen, S., Liu, F., Chen, Z., Duan, Y., Zhang, J., Wang, Y.: Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928 (2024)

    2024 arXiv

  38. [46]

    arXiv preprint arXiv:2503.05638 (2025)

    YU, M., Hu, W., Xing, J., Shan, Y.: Trajec- torycrafter: Redirecting camera trajectory for monocular videos via diffusion models. arXiv preprint arXiv:2503.05638 (2025)

    2025 arXiv

  39. [47]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)

    2022

  40. [48]

    arXiv preprint arXiv:2503.20314 (2025)

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.-W., Chen, D., Yu, F., Zhao, H., Yang, J., Zeng, J., Wang, J., Zhang, J., Zhou, J., Wang, J., Chen, J., Zhu, K., Zhao, K., Yan, K., Huang, L., Feng, M., Zhang, N., Li, P., Wu, P., Chu, R., Feng, R., Zhang, S., Sun, S., Fang, ...

    2025 arXiv

  41. [49]

    arXiv preprint arXiv:2503.16396 (2025)

    Yao, C.-H., Xie, Y., Voleti, V., Jiang, H., Jampani, V.: Sv4d 2.0: Enhancing spatio- temporal consistency in multi-view video dif- fusion for high-quality 4d generation. arXiv preprint arXiv:2503.16396 (2025)

    2025 arXiv

  42. [50]

    In: International Conference on Learning Representations (2022)

    Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large lan- guage models. In: International Conference on Learning Representations (2022)

    2022

  43. [51]

    In: European Conference on Computer Vision (2025)

    Voleti, V., Yao, C.-H., Boss, M., Letts, A., Pankratz, D., Tochilkin, D., Laforte, C., Rombach, R., Jampani, V.: Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In: European Conference on Computer Vision (2025)

    2025

  44. [52]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image dif- fusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)

    2023

  45. [53]

    In: Pro- ceedings of the AAAI Conference on Artificial Intelligence (2024)

    Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In: Pro- ceedings of the AAAI Conference on Artificial Intelligence (2024)

    2024

  46. [54]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    Hu, L.: Animate anyone: Consistent and con- trollable image-to-video synthesis for char- acter animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    2024

  47. [55]

    In: Advances in Neural Information Processing Systems (2024)

    Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., Xu, J., Chen, K., Xu, B., Li, J., Dong, Y., Ding, M., Tang, J.: Cogvlm: Visual expert for pretrained language models. In: Advances in Neural Information Processing Systems (2024)

    2024

  48. [56]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Springer Nature 2022 LATEX template 18 Article Title (2018)

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effective- ness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Springer Nature 2022 LATEX template 18 Article Title (2018)

    2022

  49. [57]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)

    Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)

    2023

  50. [58]

    In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (2023)

    QI, C., Cun, X., Zhang, Y., Lei, C., Wang, X., Shan, Y., Chen, Q.: Fatezero: Fusing atten- tions for zero-shot text-based video editing. In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (2023)

    2023

  51. [59]

    arXiv preprint arXiv:2205.15868 (2022)

    Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868 (2022)

    2022 arXiv

  52. [60]

    arXiv preprint arXiv:2408.06072 (2025)

    Yang, Z., Teng, J., Zheng, W., Ding, M., Huang, S., Xu, J., Yang, Y., Hong, W., Zhang, X., Feng, G., Yin, D., Zhang, Y., Wang, W., Cheng, Y., Xu, B., Gu, X., Dong, Y., Tang, J.: Cogvideox: Text-to-video dif- fusion models with an expert transformer. arXiv preprint arXiv:2408.0...

    2025 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.