Pith. sign in

REVIEW 3 major objections 5 minor 147 references

This paper claims that dynamic 360-degree human assets can be generated directly from text prompts as structured 4D Gaussian splats, skipping the video-first, reconstruction-later pipeline and producing consistent results in under one minut

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 04:19 UTC pith:EB6XAJ3I

load-bearing objection Direct text-to-4DGS generation for humans is a real contribution, but the main consistency comparison mixes raw videos with rendered 4D assets and needs to be made apples-to-apples. the 3 major comments →

arxiv 2607.27634 v1 pith:EB6XAJ3I submitted 2026-07-30 cs.CV

4DHumanDiff: Direct Text-to-4DGS Generation for Consistent 360-Degree Dynamic Humans

Renlong Wu , Haoran Chen , Yuxiang Wei , Xiaowei Jin , Wangmeng Zuo , Hui Li This is my paper
classification cs.CV
keywords text-to-4D generation4D Gaussian Splattingdynamic human generation360-degree consistencytemporal consistencydiffusion modelsstructured voxel representationhuman assets
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

4DHumanDiff's claim is that the bottleneck in text-to-4D human generation is the intermediate video: generating videos first and then reconstructing a 4D representation is slow, view-inconsistent, and geometrically incomplete. The paper proposes to instead model a structured 4D Gaussian Splatting space directly with a diffusion model, so that a text prompt denoises into a temporally bound sequence of voxel grids that render from any viewpoint. To make this possible, it converts reconstructed dynamic humans into a fixed-size, voxelized 4DGS representation, builds a 60,000-pair text-to-4DGS training set, and adds temporal attention plus 2D rendering supervision. The reported payoff is a one-minute generation pipeline that beats two-stage baselines on temporal and multi-view consistency and runs more than 10x faster.

Core claim

The paper's central claim is that text-to-4D generation for humans should be formulated directly in a structured 4D Gaussian Splatting space, not through an intermediate video. To make this possible, it converts reconstructed dynamic humans into a voxel-grid representation with a fixed number of Gaussians, normalizes scale, and—crucially—assigns every later timestamp to the same grid positions as the first via temporal binding. A 3D U-Net with temporal attention then denoises a noisy 4D sample conditioned on text, supervised by both a 4D reconstruction loss and a 2D rendering loss. The authors report that the resulting model renders consistently from any viewpoint, improves temporal and mult

What carries the argument

The load-bearing object is a structured 4DGS tensor: a sequence of timestamps, each a D×D×D voxel grid whose cells hold Gaussian parameters (position, scale, rotation, opacity, color). The paper obtains it by fitting dynamic humans with a fixed Gaussian count, normalizing to a unit cube, and using Optimal Transport to voxelize only the first timestamp; later timestamps inherit the same grid positions through temporal binding, so the same cell across time tracks the same surface point. This turns an unordered Gaussian cloud into a regular tensor that a standard 3D U-Net can diffuse, with temporal attention layers added to model motion and with color/opacity shared across time to enforce appea

Load-bearing premise

The fixed voxel-grid correspondence computed at the first timestamp remains valid for all later timestamps; if Gaussian trajectories cross or topology changes, the structured representation and the learned 4D distribution inherit scrambled correspondences.

What would settle it

Generate a prompt involving a large crossing motion (e.g., a dancer crossing both arms over the chest) and render the output from multiple viewpoints; if per-Gaussian identity within a voxel cell is maintained only at the first timestamp, the rendered limb will flicker, split, or ghost as Gaussians swap order—a concrete failure of temporal binding that would undermine the consistency claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A text prompt can directly yield a 4D Gaussian asset with controllable camera viewpoints, without any video generation or per-scene fitting step.
  • Temporal and multi-view consistency improve relative to video-first pipelines because consistency is baked into the representation rather than enforced after the fact.
  • Inference time drops below one minute, more than 10x faster than the compared two-stage methods, making interactive text-to-4D asset creation plausible.
  • Motion smoothness can be increased by interpolating between temporally aligned Gaussian states with no additional training.
  • The 60,000-pair dataset of structured text-to-4DGS samples makes training from scratch possible, removing dependence on pretrained video diffusion priors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if temporal binding holds, the same voxel-grid trick likely transfers to other articulated or deformable objects, since it only requires a consistent correspondence across time.
  • Editorial inference: the fixed grid correspondence is the fragile point—prompts involving limb crossing, loose clothing, or self-contact may scramble the correspondence, showing up as flicker or ghosting in rendered video.
  • Editorial inference: the one-minute ceiling is probably conservative if the model is distilled or run on larger batches, and the tensor form opens the door to latent-space diffusion at higher resolution.
  • Editorial inference: a sharper test of the central claim would compare against video-first pipelines on a held-out set of 360-degree motion prompts using an objective per-viewpoint consistency metric, rather than a user preference study.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes 4DHumanDiff, a diffusion framework that directly generates structured 4D Gaussian Splatting (4DGS) representations of dynamic humans from text prompts, bypassing the usual two-stage pipeline of text-to-video generation followed by per-scene 4D reconstruction. To make 4DGS amenable to diffusion, the authors build a large text-to-4DGS dataset from MVHumanNet by fitting GauHuman with a fixed Gaussian count, normalizing to a unit cube, and voxelizing the first timestamp via Optimal Transport while binding later timestamps to the same grid positions. The generation model is a 3D U-Net with temporal attention, trained with a 4D reconstruction loss and a 2D rendered-view regularization loss, plus a training-free 4D interpolation step for motion smoothness. The paper claims state-of-the-art temporal and multi-view consistency, and reports more than 10x faster inference than two-stage baselines, supported by quantitative comparisons, ablations, and a user study.

Significance. If the central claims hold, 4DHumanDiff is a meaningful step forward: it demonstrates that a structured 4DGS representation can be generated end-to-end from text, avoiding the computational cost and view-inconsistency problems of video-first pipelines. The construction of a 60,000-sample text-to-4DGS dataset is a valuable resource, and the ablations cover the main design choices, including temporal attention, loss terms, view count, and temporal voxel binding. However, the quantitative evidence for the headline superiority is weakened by a comparison protocol that evaluates baseline multi-view videos in raw form while evaluating 4DHumanDiff on rendered 4DGS, and by the absence of error bars or significance tests for the small margins in several ablation tables. These issues are fixable and do not invalidate the core idea, but they currently prevent full confidence in the stated conclusions.

major comments (3)
  1. [§4.3, Table 1] The central consistency comparison is not apples-to-apples. For GenXD, SV4D 2.0, and 4Diffusion, metrics are computed directly on the generated multi-view videos, whereas for 4DHumanDiff they are computed on videos rendered from the generated 4DGS. Rendering from a shared 3D/4D model can systematically raise DINO-based multi-view consistency, because all views are re-projections of a single coherent representation even when per-view sharpness is lower. The paper's justification that this protocol 'excludes reconstruction degradation' assumes the effect is monotonic in favor of the baselines; that assumption is not established. The 0.862 vs 0.798 margin in Table 1 is exactly the quantity that a full-pipeline comparison must test. Please evaluate the baselines after 4D reconstruction (reporting success/failure rates), or evaluate all methods by rendering their outputs at identical camera t
  2. [Tables 1-7] All quantitative claims rest on a single evaluation run with no error bars or significance tests. In Tables 4, 5, and 7 several differences are 0.001-0.02 metric points; for example, Multi-View Consistency is 0.862 in both the w/oExc. and Ours rows of Table 4 and in both w/oTem. and Ours in Table 5, and Subject Consistency changes from 0.973 to 0.984 in Table 7. With no variance estimates, the conclusions in §5.1-5.4 that each component improves performance are not supported. Please report at least three seeds or bootstrap confidence intervals and a paired significance test for the main comparisons, especially where the reported margins are tiny.
  3. [§3.2, §5.4] The temporal voxel binding ablation compares against independent per-timestamp voxelization (w/o Binding), which conflates two effects: loss of temporal correspondence and change of spatial layout across timestamps. This makes it difficult to attribute the Table 7 gains specifically to correspondence preservation. To isolate the proposed mechanism, add an intermediate baseline that retains a single OT ordering but randomly permutes Gaussian identities at a subset of timestamps, breaking only the temporal binding. This is important because §5.4 is the only direct evidence for the key structural premise of the representation.
minor comments (5)
  1. [Eq. (10)] The symbol \(\odot\) is used but not defined. If the interpolation is applied to quaternion rotations and scales, direct linear interpolation of quaternions can produce non-unit quaternions; please specify normalization or state how the representation channels are treated.
  2. [Table 1] The 'Inference Time' column needs a precise definition. Does the reported time for two-stage baselines include both video generation and 4D reconstruction, or only the reconstruction stage? For L4GM, does it include the time of the preceding Wan 2.1 generation? Without this, the 'over 10x faster' claim is difficult to verify.
  3. [§4.1] The normalization with dataset statistics \(\bar\mu\) and \(\bar\sigma\) is mentioned but the exact forward and inverse formulas are not given. Please include them, since the diffusion model operates on normalized 4DGS and the final output must be unnormalized for rendering.
  4. [User Study] The user study reports preference rates of 64.8% and 76.4% but gives no number of paired comparisons per participant, no confidence intervals, and no test of significance. Please add these details.
  5. [General] There are minor typographical issues, e.g., 'T able' in Table captions and 'RESEARCH P APER' in the header. Also, references [77] and [131] appear to be the same paper; please deduplicate.

Circularity Check

0 steps flagged

No circularity identified; the pipeline is trained on externally fitted data and evaluated with external metrics, and no prediction reduces to its inputs by construction.

full rationale

The paper's derivation chain is self-contained in the relevant sense. The 4DGS training targets G0 are externally fitted GauHuman reconstructions from MVHumanNet (Sec. 3.2), and the diffusion training loss L4d = ||D(Gt,Y)-G0||_1 (Eq. 6) is a standard reconstruction objective; no fitted parameter is later renamed as a prediction. The temporal voxel binding is a dataset/representation design choice: OT voxelization is computed once at the first timestamp and propagated through the GauHuman temporal correspondence, and the 4D interpolation of Eq. (10) is a direct linear blend of this aligned representation, explicitly training-free and not used as evidence for the headline consistency claim. Evaluation uses external quantities (MUSIQ, DINO, AMT, ViCLIP, and a user study) that are not optimization targets of the model. The two self-citations ([35], [132]) concern video interpolation and a 4DGS deformation assumption; they are not load-bearing for the central text-to-4DGS result, and the paper invokes no uniqueness theorem or prior-work ansatz to force its representation choice. The Table 1 protocol asymmetry (raw multi-view video baselines vs rendered 4DGS for the proposed method) is a legitimate benchmarking caveat, but it is not a reduction-by-construction of the reported consistency metrics. The Limitations section (Sec. 6) only notes missing local-detail refinement, not a circular step. Therefore, no circularity is present.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central claim rests primarily on the fixed-count voxelized 4DGS representation, the temporal-binding correspondence, and a set of hand-chosen hyperparameters and dataset assumptions. No new physical entities are introduced. The most consequential unverified premises are the quality of the 60k fitted 4DGS dataset and the validity of temporal voxel binding under large deformation.

free parameters (7)
  • N_max (fixed Gaussian count) = 13824
    Chosen capacity for the structured 4DGS grid; all clips are fitted to this count using densification-constrained fitting (Section 3.2).
  • D (voxel grid resolution) = 24
    Cube root of N_max; sets the spatial layout resolution of the voxelized 4DGS (Sections 3.2, 4.1).
  • C (per-Gaussian feature channels) = 14
    x(3)+r(4)+s(3)+alpha(1)+c(3) per Gaussian; representation design choice (Section 4.1).
  • lambda (2D regularization weight) = 10
    Loss weight in Eq. (5); chosen by hand to balance 4D reconstruction and 2D rendering losses (Section 4.1).
  • M (number of rendered views for L2d) = 4
    Views rendered for 2D regularization; ablation shows M=4 and M=8 are nearly identical (Section 5.1).
  • W (4D interpolation factor) = 8
    Number of interpolated states in Eq. (10); user-set temporal smoothness factor (Sections 4.1, 5.3).
  • mu-bar, sigma-bar (dataset normalization statistics) = computed from 60k dataset
    Per-channel mean/std used to normalize training samples (Section 4.1); fitted statistics rather than tuned parameters.
axioms (6)
  • domain assumption Optimal transport can assign 13,824 Gaussians to a 24^3 voxel grid with a meaningful spatial layout.
    Section 3.2, third step, relies on OT voxelization to establish a fixed spatial index; if the OT layout is not semantically consistent across samples, the diffusion target is noisy.
  • domain assumption Human dynamic appearance can be represented by 4DGS with a fixed Gaussian count, where only positions and scales vary over time and color/opacity are shared.
    Section 3.3 uses this common prior from 4D reconstruction works [53-55, 62, 63, 131, 132] to define the structured target; failure of this prior for loose clothing or topology change would degrade outputs.
  • domain assumption The 60,000 fitted 4DGS samples from MVHumanNet are diverse and accurate enough to train a generalizable text-to-4D model.
    Section 3.2 selects MVHumanNet and Gemini-generated captions; no distribution analysis or failure-rate audit is provided.
  • domain assumption Gemini 1.5 text annotations and evaluation prompts are semantically accurate and unbiased for human appearance and motion.
    Sections 3.2 and 4.2 rely on Gemini for both training captions and the 170 evaluation prompts; annotation errors would propagate into model training and evaluation.
  • domain assumption Gradients from rendering-based 2D losses (L1, VGG) through 4DGS are stable and useful for training.
    Section 3.3, Eqs. (7)-(9), use rendered videos as supervision; this assumes differentiable rendering and patch-selection stability.
  • domain assumption The structured-Gaussian diffusion formulation (GaussianCube-style) scales to the 4D temporal axis without losing spatial structure.
    Section 3.3 extends a 3D U-Net to 4D with temporal attention; there is no theoretical guarantee that temporal attention preserves Gaussian identity, so this is an empirical modeling assumption.

pith-pipeline@v1.3.0-daily-deepseek · 16406 in / 14842 out tokens · 140666 ms · 2026-08-01T04:19:17.321592+00:00 · methodology

0 comments
read the original abstract

Generating high-quality 360-degree dynamic human assets from text prompts is challenging. Existing methods usually synthesize monocular or multi-view videos first and then fit a 4D representation, which is expensive and often causes incomplete geometry or view-inconsistent renderings. We present 4DHumanDiff, a diffusion framework that directly generates dynamic humans represented by 4D Gaussian Splatting (4DGS) from text prompts. By modeling the structured 4D representation space end-to-end, 4DHumanDiff avoids video pre-generation and per-scene reconstruction, making it better suited for view-consistent and temporally coherent asset generation. The model uses a 3D U-Net backbone with temporal attention for motion-aware generation. We further construct a large-scale text-to-4DGS dataset with 60,000 high-quality pairs, and introduce 2D regularization and training-free 4D interpolation to improve rendering quality and motion smoothness. Experiments show that 4DHumanDiff generates consistent 360-degree dynamic humans within one minute, achieves better temporal and multi-view consistency, and reduces inference time by more than 10x.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

147 extracted references · 58 linked inside Pith

  1. [1]

    HumanSD : A native skeleton-guided diffusion model for human image generation

    Ju X, Zeng A, Zhao C, et al. HumanSD : A native skeleton-guided diffusion model for human image generation. In: Proceedings of IEEE/CVF International Conference on Computer Vision, 2023. 15988--15998

    2023

  2. [2]

    Self-supervised collision handling via generative 3d garment models for virtual try-on

    Santesteban I, Thuerey N, Otaduy M A, et al. Self-supervised collision handling via generative 3d garment models for virtual try-on. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 11763--11773

    2021

  3. [3]

    Ulnef: Untangled layered neural fields for mix-and-match virtual try-on

    Santesteban I, Otaduy M, Thuerey N, et al. Ulnef: Untangled layered neural fields for mix-and-match virtual try-on. Advances in Neural Information Processing Systems, 2022, 35: 12110--12125

    2022

  4. [4]

    Mv-vton: Multi-view virtual try-on with diffusion models

    Wang H, Zhang Z, Di D, et al. Mv-vton: Multi-view virtual try-on with diffusion models. In: Proceedings of AAAI Conference on Artificial Intelligence, 2025. 7682--7690

    2025

  5. [5]

    Direct inversion: Boosting diffusion-based editing with 3 lines of code

    Ju X, Zeng A, Bian Y, et al. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    Human-art: A versatile human-centric dataset bridging natural and artificial scenes

    Ju X, Zeng A, Wang J, et al. Human-art: A versatile human-centric dataset bridging natural and artificial scenes. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 618--629

    2023

  7. [7]

    Audio-driven co-speech gesture video generation

    Liu X, Wu Q, Zhou H, et al. Audio-driven co-speech gesture video generation. Advances in Neural Information Processing Systems, 2022, 35: 21386--21399

    2022

  8. [8]

    Taming diffusion models for audio-driven co-speech gesture generation

    Zhu L, Liu X, Liu X, et al. Taming diffusion models for audio-driven co-speech gesture generation. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 10544--10553

    2023

  9. [9]

    L4GM : Large 4d gaussian reconstruction model

    Ren J, Xie C, Mirzaei A, et al. L4GM : Large 4d gaussian reconstruction model. Advances in Neural Information Processing Systems, 2024, 37: 56828--56858

    2024

  10. [10]

    Voyager: Long-range and world-consistent video diffusion for explorable 3D scene generation

    Huang T, Zheng W, Wang T, et al. Voyager: Long-range and world-consistent video diffusion for explorable 3D scene generation. arXiv preprint arXiv:2506.04225, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Adding conditional control to text-to-image diffusion models

    Zhang L, Rao A, Agrawala M. Adding conditional control to text-to-image diffusion models. In: Proceedings of IEEE/CVF International Conference on Computer Vision, 2023. 3836--3847

    2023

  12. [12]

    CogVideoX : Text-to-video diffusion models with an expert transformer

    Yang Z, Teng J, Zheng W, et al. CogVideoX : Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Wan: Open and advanced large-scale video generative models

    Wan T, Wang A, Ai B, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    GenXD : Generating any 3d and 4d scenes

    Zhao Y, Lin C C, Lin K, et al. GenXD : Generating any 3d and 4d scenes. arXiv preprint arXiv:2411.02319, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    SV4D 2.0: Enhancing spatio-temporal consistency in multi-view video diffusion for high-quality 4d generation

    Yao C H, Xie Y, Voleti V, et al. SV4D 2.0: Enhancing spatio-temporal consistency in multi-view video diffusion for high-quality 4d generation. arXiv preprint arXiv:2503.16396, 2025

    Pith/arXiv arXiv 2025

  16. [16]

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

    Zhang H, Chen X, Wang Y, et al. 4Diffusion : Multi-view video diffusion model for 4d generation. Advances in Neural Information Processing Systems, 2024, 37: 15272--15295

    2024

  17. [17]

    GauHuman : Articulated gaussian splatting from monocular human videos

    Hu S, Hu T, Liu Z. GauHuman : Articulated gaussian splatting from monocular human videos. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 20418--20431

    2024

  18. [18]

    Real-time high-fidelity gaussian human avatars with position-based interpolation of spatially distributed mlps

    Zhan Y, Shao T, Yang Y, et al. Real-time high-fidelity gaussian human avatars with position-based interpolation of spatially distributed mlps. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 26297--26307

    2025

  19. [19]

    3DGS -avatar: Animatable avatars via deformable 3D gaussian splatting

    Qian Z, Wang S, Mihajlovic M, et al. 3DGS -avatar: Animatable avatars via deformable 3D gaussian splatting. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5020--5030

    2024

  20. [20]

    Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling

    Li Z, Zheng Z, Wang L, et al. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 19711--19722

    2024

  21. [21]

    R3-avatar: Record and retrieve temporal codebook for reconstructing photorealistic human avatars

    Zhan Y, Xu W, Zhu Q, et al. R3-avatar: Record and retrieve temporal codebook for reconstructing photorealistic human avatars. arXiv preprint arXiv:2503.12751, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    4D gaussian splatting for real-time dynamic scene rendering

    Wu G, Yi T, Fang J, et al. 4D gaussian splatting for real-time dynamic scene rendering. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 20310--20320

    2024

  23. [23]

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

    Yang Z, Yang H, Pan Z, et al. Real-time photorealistic dynamic scene representation and rendering with 4D gaussian splatting. arXiv preprint arXiv:2310.10642, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Improved denoising diffusion probabilistic models

    Nichol A Q, Dhariwal P. Improved denoising diffusion probabilistic models. In: Proceedings of International Conference on Machine Learning. PMLR, 2021. 8162--8171

    2021

  25. [25]

    Diffusion models beat gans on image synthesis

    Dhariwal P, Nichol A. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 2021, 34: 8780--8794

    2021

  26. [26]

    GaussianCube : Structuring gaussian splatting using optimal transport for 3d generative modeling

    Zhang B, Cheng Y, Yang J, et al. GaussianCube : Structuring gaussian splatting using optimal transport for 3d generative modeling. CoRR, 2024

    2024

  27. [27]

    MVHumanNet : A large-scale dataset of multi-view daily dressing human captures

    Xiong Z, Li C, Liu K, et al. MVHumanNet : A large-scale dataset of multi-view daily dressing human captures. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 19801--19811

    2024

  28. [28]

    NeRF : Representing scenes as neural radiance fields for view synthesis

    Mildenhall B, Srinivasan P P, Tancik M, et al. NeRF : Representing scenes as neural radiance fields for view synthesis. Commun. ACM, 2021, 65: 99--106

    2021

  29. [29]

    Mip- NeRF : A multiscale representation for anti-aliasing neural radiance fields

    Barron J T, Mildenhall B, Tancik M, et al. Mip- NeRF : A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of IEEE/CVF International Conference on Computer Vision, 2021. 5855--5864

    2021

  30. [30]

    NeRF in the wild: Neural radiance fields for unconstrained photo collections

    Martin-Brualla R, Radwan N, Sajjadi M S, et al. NeRF in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 7210--7219

    2021

  31. [31]

    Free NeRF : Improving few-shot neural rendering with free frequency regularization

    Yang J, Pavone M, Wang Y. Free NeRF : Improving few-shot neural rendering with free frequency regularization. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 8254--8263

    2023

  32. [32]

    D- NeRF : Neural radiance fields for dynamic scenes

    Pumarola A, Corona E, Pons-Moll G, et al. D- NeRF : Neural radiance fields for dynamic scenes. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 10318--10327

    2021

  33. [33]

    Instant neural graphics primitives with a multiresolution hash encoding

    M \"u ller T, Evans A, Schied C, et al. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022, 41: 1--15

    2022

  34. [34]

    3D gaussian splatting for real-time radiance field rendering

    Kerbl B, Kopanas G, Leimk \"u hler T, et al. 3D gaussian splatting for real-time radiance field rendering. ACM, 2023, 42: 139--1

    2023

  35. [35]

    Dual-camera smooth zoom on mobile phones

    Wu R, Zhang Z, Yang Y, et al. Dual-camera smooth zoom on mobile phones. In: Proceedings of European Conference on Computer Vision. Springer, 2024. 250--269

    2024

  36. [36]

    Colmap-free 3D gaussian splatting

    Fu Y, Liu S, Kulkarni A, et al. Colmap-free 3D gaussian splatting. arXiv preprint arXiv:2312.07504, 2023

    Pith/arXiv arXiv 2023

  37. [37]

    Depth-regularized optimization for 3D gaussian splatting in few-shot images

    Chung J, Oh J, Lee K M. Depth-regularized optimization for 3D gaussian splatting in few-shot images. arXiv preprint arXiv:2311.13398, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    Cosseggaussians: Compact and swift scene segmenting 3D gaussians

    Dou B, Zhang T, Ma Y, et al. Cosseggaussians: Compact and swift scene segmenting 3D gaussians. arXiv preprint arXiv:2401.05925, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Tip-editor: An accurate 3D editor following both text-prompts and image-prompts

    Zhuang J, Kang D, Cao Y P, et al. Tip-editor: An accurate 3D editor following both text-prompts and image-prompts. arXiv preprint arXiv:2401.14828, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    Mip-splatting: Alias-free 3D gaussian splatting

    Yu Z, Chen A, Huang B, et al. Mip-splatting: Alias-free 3D gaussian splatting. arXiv preprint arXiv:2311.16493, 2023

    Pith/arXiv arXiv 2023

  41. [41]

    Compact 3d scene representation via self-organizing gaussian grids

    Morgenstern W, Barthel F, Hilsmann A, et al. Compact 3d scene representation via self-organizing gaussian grids. arXiv preprint arXiv:2312.13299, 2023

    Pith/arXiv arXiv 2023

  42. [42]

    Ln3Diff : Scalable latent neural fields diffusion for speedy 3d generation

    Lan Y, Hong F, Yang S, et al. Ln3Diff : Scalable latent neural fields diffusion for speedy 3d generation. In: Proceedings of European Conference on Computer Vision. Springer, 2024. 112--130

    2024

  43. [43]

    DiffGS : Functional gaussian splatting diffusion

    Zhou J, Zhang W, Liu Y S. DiffGS : Functional gaussian splatting diffusion. Advances in Neural Information Processing Systems, 2024, 37: 37535--37560

    2024

  44. [44]

    L3DG : Latent 3d gaussian diffusion

    Roessle B, M \"u ller N, Porzi L, et al. L3DG : Latent 3d gaussian diffusion. In: Proceedings of SIGGRAPH Asia 2024 Conference Papers, 2024. 1--11

    2024

  45. [45]

    Baking gaussian splatting into diffusion denoiser for fast and scalable single-stage image-to-3d generation and reconstruction

    Cai Y, Zhang H, Zhang K, et al. Baking gaussian splatting into diffusion denoiser for fast and scalable single-stage image-to-3d generation and reconstruction. In: Proceedings of IEEE/CVF International Conference on Computer Vision, 2025. 25062--25072

    2025

  46. [46]

    Novelgs: Consistent novel-view denoising via large gaussian reconstruction model

    Liu J, Xu J, Cheng W, et al. Novelgs: Consistent novel-view denoising via large gaussian reconstruction model. arXiv preprint arXiv:2411.16779, 2024

    Pith/arXiv arXiv 2024

  47. [47]

    VideoRFSplat : Direct scene-level text-to- 3D gaussian splatting generation with flexible pose and multi-view joint modeling

    Go H, Park B, Nam H, et al. VideoRFSplat : Direct scene-level text-to- 3D gaussian splatting generation with flexible pose and multi-view joint modeling. arXiv preprint arXiv:2503.15855, 2025

    Pith/arXiv arXiv 2025

  48. [48]

    Lgm: Large multi-view gaussian model for high-resolution 3d content creation

    Tang J, Chen Z, Chen X, et al. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In: Proceedings of European Conference on Computer Vision. Springer, 2024. 1--18

    2024

  49. [49]

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

    Fridovich-Keil S, Meanti G, Warburg F R, et al. K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 12479--12488

    2023

  50. [50]

    Forward flow for novel view synthesis of dynamic scenes

    Guo X, Sun J, Dai Y, et al. Forward flow for novel view synthesis of dynamic scenes. In: Proceedings of IEEE/CVF International Conference on Computer Vision, 2023. 16022--16033

    2023

  51. [51]

    3d motion magnification: Visualizing subtle motions from time-varying radiance fields

    Feng B Y, Alzayer H, Rubinstein M, et al. 3d motion magnification: Visualizing subtle motions from time-varying radiance fields. In: Proceedings of IEEE/CVF International Conference on Computer Vision, 2023. 9837--9846

    2023

  52. [52]

    Factorized motion fields for fast sparse input dynamic view synthesis

    Somraj N, Choudhary K, Mupparaju S H, et al. Factorized motion fields for fast sparse input dynamic view synthesis. In: Proceedings of ACM SIGGRAPH 2024 Conference Papers, 2024. 1--12

    2024

  53. [53]

    4D gaussian splatting: Towards efficient novel view synthesis for dynamic scenes

    Duan Y, Wei F, Dai Q, et al. 4D gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. arXiv, 2024

    2024

  54. [54]

    3d geometry-aware deformable gaussian splatting for dynamic view synthesis

    Lu Z, Guo X, Hui L, et al. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In: Proceedings of CVPR, 2024. 8900--8910

    2024

  55. [55]

    Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis

    Luiten J, Kopanas G, Leibe B, et al. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv, 2023

    2023

  56. [56]

    Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3D gaussian splatting

    Kratimenos A, Lei J, Daniilidis K. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3D gaussian splatting. arXiv preprint arXiv:2312.00112, 2023

    Pith/arXiv arXiv 2023

  57. [57]

    Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes

    Huang Y H, Sun Y T, Yang Z, et al. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4220--4230

    2024

  58. [58]

    Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle

    Lin Y, Dai Z, Zhu S, et al. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 21136--21145

    2024

  59. [59]

    Spacetime gaussian feature splatting for real-time dynamic view synthesis

    Li Z, Chen Z, Li Z, et al. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 8508--8520

    2024

  60. [60]

    Swings: Sliding windows for dynamic 3D gaussian splatting

    Shaw R, Nazarczuk M, Song J, et al. Swings: Sliding windows for dynamic 3D gaussian splatting

  61. [61]

    3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos

    Sun J, Jiao H, Li G, et al. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 20675--20685

    2024

  62. [62]

    4D gaussian splatting for real-time dynamic scene rendering

    Wu G, Yi T, Fang J, et al. 4D gaussian splatting for real-time dynamic scene rendering. In: Proceedings of CVPR, 2024. 20310--20320

    2024

  63. [63]

    Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction

    Yang Z, Gao X, Zhou W, et al. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In: Proceedings of CVPR, 2024. 20331--20341

    2024

  64. [64]

    Per-gaussian embedding-based deformation for deformable 3D gaussian splatting

    Bae J, Kim S, Yun Y, et al. Per-gaussian embedding-based deformation for deformable 3D gaussian splatting. arXiv preprint arXiv:2404.03613, 2024

    Pith/arXiv arXiv 2024

  65. [65]

    Splatfields: Neural gaussian splats for sparse 3D and 4D reconstruction

    Mihajlovic M, Prokudin S, Tang S, et al. Splatfields: Neural gaussian splats for sparse 3D and 4D reconstruction. arXiv preprint arXiv:2409.11211, 2024

    Pith/arXiv arXiv 2024

  66. [66]

    4D -rotor gaussian splatting: towards efficient novel view synthesis for dynamic scenes

    Duan Y, Wei F, Dai Q, et al. 4D -rotor gaussian splatting: towards efficient novel view synthesis for dynamic scenes. In: Proceedings of ACM SIGGRAPH 2024 Conference Papers, 2024. 1--11

    2024

  67. [67]

    Dreamscene4d: Dynamic multi-object scene generation from monocular videos

    Chu W H, Ke L, Fragkiadaki K. Dreamscene4d: Dynamic multi-object scene generation from monocular videos. arXiv preprint arXiv:2405.02280, 2024

    Pith/arXiv arXiv 2024

  68. [68]

    A compact dynamic 3D gaussian representation for real-time dynamic view synthesis

    Katsumata K, Vo D M, Nakayama H. A compact dynamic 3D gaussian representation for real-time dynamic view synthesis. In: Proceedings of ECCV, 2024

    2024

  69. [69]

    Fast view synthesis of casual videos

    Lee Y C, Zhang Z, Blackburn-Matzen K, et al. Fast view synthesis of casual videos. arXiv preprint arXiv:2312.02135, 2023

    Pith/arXiv arXiv 2023

  70. [70]

    4D gaussian splatting for high-fidelity dynamic reconstruction of single-view scenes

    Yang X, Xie W, Fu Y, et al. 4D gaussian splatting for high-fidelity dynamic reconstruction of single-view scenes. SSRN 4895336

  71. [71]

    Mono NeRF : Learning a generalizable dynamic radiance field from monocular videos

    Tian F, Du S, Duan Y. Mono NeRF : Learning a generalizable dynamic radiance field from monocular videos. In: Proceedings of IEEE/CVF International Conference on Computer Vision, 2023. 17903--17913

    2023

  72. [72]

    Gaussianflow: Splatting gaussian dynamics for 4d content creation

    Gao Q, Xu Q, Cao Z, et al. Gaussianflow: Splatting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024

    Pith/arXiv arXiv 2024

  73. [73]

    Motion-aware 3D gaussian splatting for efficient dynamic scene reconstruction

    Guo Z, Zhou W, Li L, et al. Motion-aware 3D gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024

    Pith/arXiv arXiv 2024

  74. [74]

    Modgs: Dynamic gaussian splatting from casually captured monocular videos

    Liu Q, Liu Y, Wang J, et al. Modgs: Dynamic gaussian splatting from casually captured monocular videos. arXiv preprint arXiv:2406.00434, 2024

    Pith/arXiv arXiv 2024

  75. [75]

    Motiongs: Exploring explicit motion guidance for deformable 3D gaussian splatting

    Zhu R, Liang Y, Chang H, et al. Motiongs: Exploring explicit motion guidance for deformable 3D gaussian splatting. arXiv preprint arXiv:2410.07707, 2024

    Pith/arXiv arXiv 2024

  76. [76]

    Gflow: Recovering 4D world from monocular video

    Wang S, Yang X, Shen Q, et al. Gflow: Recovering 4D world from monocular video. arXiv preprint arXiv:2405.18426, 2024

    Pith/arXiv arXiv 2024

  77. [77]

    Shape of motion: 4d reconstruction from a single video

    Wang Q, Ye V, Gao H, et al. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024

    arXiv 2024

  78. [78]

    Dynomo: Online point tracking by dynamic online monocular gaussian reconstruction

    Seidenschwarz J, Zhou Q, Duisterhof B, et al. Dynomo: Online point tracking by dynamic online monocular gaussian reconstruction. arXiv preprint arXiv:2409.02104, 2024

    Pith/arXiv arXiv 2024

  79. [79]

    Dynamic gaussian marbles for novel view synthesis of casual monocular videos

    Stearns C, Harley A, Uy M, et al. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. arXiv preprint arXiv:2406.18717, 2024

    Pith/arXiv arXiv 2024

  80. [80]

    Mosca: Dynamic gaussian fusion from casual videos via 4D motion scaffolds

    Lei J, Weng Y, Harley A, et al. Mosca: Dynamic gaussian fusion from casual videos via 4D motion scaffolds. arXiv preprint arXiv:2405.17421, 2024

    Pith/arXiv arXiv 2024

Showing first 80 references.