Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Text-driven mesh animation can be learned from raw mesh sequences by recovering pseudo-skeletons once, then generating editable skeleton motion without per-frame labels.

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 · grok-4.5

2026-07-10 10:41 UTC pith:OEZNZRKH

load-bearing objection Solid systems paper: weakly supervised pseudo-skeletons + feed-forward text-to-skeleton motion that is competitive with fully supervised baselines and more editable than SDS/video deformation. the 3 major comments →

arxiv 2607.08246 v1 pith:OEZNZRKH submitted 2026-07-09 cs.CV

SkelGen4D: Weakly-Supervised Skeleton-Based 4D Generation for Text-Driven Mesh Animation

Hao Feng , Zhi Zuo , Jia-Hui Pan , Ka-Hei Hui , Zhengzhe Liu , Dian Zhang , Haoran Xie , Bin Sheng
show 1 more author
Jingyu Hu
This is my paper
classification cs.CV
keywords mesh animation4D generationskeleton motion learningweak supervisiontext-driven animationlinear blend skinningMotion-GRPO
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.

Industrial animation is built on skeletons that drive meshes through skinning, but most learning-based 4D systems either optimize expensive per-instance volumetric models or animate meshes with opaque deformation fields that are hard to edit. SkelGen4D argues you can stay in the industrial skeleton pipeline without needing dense per-frame skeleton annotations. It first fits a single automatic rig on the first frame and recovers temporally consistent joint transforms by matching skinned reconstructions to the observed mesh sequence. Those pseudo-skeletons then train a text-conditioned feed-forward generator, refined by a motion-level preference objective for smoothness, dynamics, and bone-length consistency. On animal motion data the weakly supervised model matches or beats fully supervised skeleton generators; on a broader 4D mesh benchmark it produces more stable, coherent animations than video-driven and score-distillation baselines, while still supporting composition, rescaling, and local editing.

Core claim

Weakly supervised skeleton modeling from raw animated meshes—without per-frame skeleton annotations—is sufficient to train a feed-forward text-to-skeleton generator that matches or surpasses fully supervised skeleton baselines and outperforms video-driven and SDS-based 4D methods, while remaining editable in standard production pipelines.

What carries the argument

Two-stage SkelGen4D: (1) differentiable LBS fitting that freezes first-frame topology and skinning weights and optimizes per-frame joint transforms so the skinned mesh tracks the ground-truth sequence, yielding pseudo-skeletons; (2) a latent auto-regressive transformer that generates text-conditioned skeleton sequences, refined by Motion-GRPO rewards for smoothness, dynamics, and bone-length consistency.

Load-bearing premise

A single automatic first-frame rig plus fixed skinning weights, with joint poses fitted only by matching skinned vertices to the mesh, produces control structures that are faithful enough to teach general text-driven motion across many object types.

What would settle it

On held-out Truebones Zoo or Diffusion4D sequences, if the Stage-1 pseudo-skeletons systematically fail to reconstruct mesh motion (high vertex error or collapsed topology) or if a fully supervised skeleton generator trained on true labels clearly beats the weakly supervised model on coverage and diversity while remaining editable, the central claim fails.

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

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. SkelGen4D proposes a two-stage, weakly supervised pipeline for text-driven mesh animation that produces explicit skeleton motions without per-frame skeleton labels. Stage 1 obtains a fixed topology and skinning weights from UniRig on the first frame, then recovers temporally consistent pseudo-skeleton poses by gradient-based fitting of joint transforms so that LBS-reconstructed meshes match the observed mesh sequence (Eqs. 1–4). Stage 2 encodes poses into a fixed latent space and trains a text-conditioned auto-regressive transformer, refined by Motion-GRPO with smoothness, dynamics, and bone-length rewards. At inference, generated skeletons drive the input mesh via LBS. On Truebones Zoo the method is reported to match or exceed fully supervised skeleton generators (MDM, SinMDM, AnyTop) on Coverage and Intra Div. Diff.; on Diffusion4D it outperforms SDS and video-driven baselines under VLM and user-study scores, while supporting motion composition, rescaling, and local editing.

Significance. If the weakly supervised pseudo-skeletons are faithful enough to serve as drop-in supervision, the work is practically significant: it aligns 4D generation with industrial skeleton pipelines, removes the need for dense per-frame annotations, and demonstrates feed-forward text-to-skeleton generation with explicit editability. Strengths include dual-benchmark evaluation, an explicit controllability section (Fig. 7), and a sequence-level RL objective (Motion-GRPO) that is a concrete, falsifiable design choice rather than pure SDS distillation. The industrial-pipeline compatibility claim is a genuine differentiator relative to deformation-field and volumetric 4D methods.

major comments (3)
  1. [§III.B, Eqs. (2)–(4); Table IV; §IV.B–C] The central claim that weakly supervised skeleton modeling matches or surpasses fully supervised baselines (Abstract; §IV.C; Table I) rests on Stage-1 pseudo-skeletons being motion-faithful control structures, not merely mesh-reconstruction soft labels. Stage 1 fixes UniRig topology and W from frame 1 and optimizes only per-frame joint transforms by vertex MSE through LBS (Eqs. 2–4). Vertex MSE does not uniquely determine joint angles or bone lengths under multi-joint skinning and residual non-rigid deformation. Table IV reports only mesh MSE for alternative optimizers; there is no joint-angle, bone-length, or hierarchy error of the recovered pseudo-skeletons against Truebones ground-truth skeletons, which are available on that benchmark. Without that comparison (or an equivalent kinematic fidelity metric), high Coverage / low Intra Div. Diff. can be explained by Stage 2 matching soft la
  2. [§IV.A.3; §IV.D; Tables II–III] On Diffusion4D the paper correctly notes that provided skeletons are noisy (Fig. 6) and therefore uses video-driven baselines. However, the VLM and user-study protocols (Tables II–III) compare heterogeneous outputs (volumetric Gaussians, meshes, skeleton-skinned meshes) with a single set of perceptual axes. It is not clear whether geometric/structural scores are confounded by representation type (e.g., explicit mesh vs. SDS blur) rather than motion quality alone, nor whether the same text prompts and camera paths are used for all methods. Please document the exact VLM prompt, scoring rubric, number of samples, and whether video-driven methods receive GT-rendered driving video of the same assets (as stated) while text-only methods do not—i.e., whether the comparison is text-to-4D vs. video-to-4D under unequal conditioning.
  3. [§III.C.3, Eqs. (15)–(24); Table V] Motion-GRPO is presented as essential for temporal coherence and physical plausibility (§III.C.3; Table V). The ablation shows modest metric shifts (Coverage 92.89→93.52; Local Div. 0.215→0.260) with overlapping stds, and the reward suite introduces multiple free parameters (λ_smooth, λ_dyn, λ_bone, δ, γ, τ, r_min, K, β) that are not sensitivity-analyzed. If GRPO is load-bearing for the “physically plausible / articulated” claim, please report reward-component ablations and at least one kinematic diagnostic (e.g., bone-length variance over time, foot-skate or interpenetration proxies) rather than only the AnyTop-style diversity/fidelity suite.
minor comments (5)
  1. [§IV.A; §IV.B] Hyperparameters for Stage-2 losses (λ1, λ2), GRPO, joint cap (96), and the reconstruction-error threshold defining the ~85% success rate are listed only partially or not at all in the main text; please add a compact table or appendix.
  2. [Fig. 4; Fig. 5] Fig. 4 and Fig. 5 are hard to read in grayscale print; consider larger frames or annotated failure modes for baselines.
  3. [§III.A; §III.D] Notation: M1 vs M0 for the rest/first frame is inconsistent between §III.A and §III.D; unify.
  4. [§II.d; §V] Related work on template-free rigging (UniRig, MagicArticulate, RigAnything) is cited but the dependence of the whole pipeline on UniRig failure modes is under-discussed in Limitations.
  5. [Table IV; Fig. 5; §IV.C] Typo/style: “Mothed” in Table IV header; “A ctionMesh” / spacing issues in Fig. 5 caption; “Truebones Zoodataset” missing space in §IV.C.

Circularity Check

1 steps flagged

No load-bearing circularity: Stage-1 LBS fits produce pseudo-labels by construction, but Stage-2 generation and external-benchmark claims are not forced by those fits.

specific steps
  1. self definitional [Section IV.B (Analysis of Auto-Rigging), success-rate definition; cf. Stage-1 Eqs. (2)–(4)]
    "A skeleton is considered successful if it satisfies two criteria: (i) it forms a valid articulated structure without topology collapse ... and (ii) it enables stable reconstruction of the input mesh sequence, measured by a reconstruction error below a predefined threshold. ... Our method achieves a success rate of approximately 85%"

    Criterion (ii) is essentially the same L_mesh vertex-MSE objective that Stage 1 already minimizes by construction when recovering P*_t. Reporting high success therefore partly restates that the optimizer found low-error solutions rather than independently validating kinematic fidelity; however this is only a robustness statistic, not the load-bearing claim that the subsequent generator matches fully-supervised baselines.

full rationale

The paper's derivation is a standard two-stage pipeline (mesh sequences → gradient-based LBS pose fitting for pseudo-skeletons → latent AR transformer + Motion-GRPO trained to match those targets and sequence rewards → LBS at inference). Evaluation of the generative claim uses held-out skeleton metrics (Coverage, diversities) against Truebones GT, VLM scores, and human studies against external baselines (AnyTop, MDM, video-driven methods). These are independent of the fitting objective. The only mild self-reference is that Stage-1 'success' is partly defined by the same vertex-MSE reconstruction used for fitting (so high success is expected when the optimizer converges); this is not used to claim the main result and does not make the weakly-vs-fully-supervised comparison tautological. No self-citation uniqueness theorems, no fitted parameter renamed as a prediction of a related quantity, and no ansatz smuggled via author-overlapping citations. The skeptic concern about non-uniqueness of joint angles under vertex MSE is a correctness/identifiability risk, not circularity under the stated criteria.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

The central claim rests on standard LBS animation math, the assumption that UniRig first-frame topology/skinning is adequate and fixed, and several hand-chosen training/reward weights. Motion-GRPO and the pseudo-skeleton fitting procedure are paper-specific constructs, not free-floating physical entities. No new particles or forces are postulated; free parameters are optimization and reward knobs that shape reported quality.

free parameters (7)
  • λ1, λ2 (latent vs pose reconstruction weights)
    Stage-2 supervised loss weights balancing L_feat and L_recon; values not specified numerically but required for the reported generator quality.
  • λ_smooth, λ_dyn, λ_bone (Motion-GRPO reward weights)
    Balance smoothness, dynamics, and bone-length rewards; chosen to produce the GRPO ablation gains.
  • δ, γ (dynamics threshold and penalty)
    Hand-set floor on motion magnitude and penalty when below it; directly shapes the non-static reward.
  • τ, r_min (bone-length deviation tolerance and clip)
    Tolerance before bone-length penalty and clipping range; control skeleton-consistency reward scale.
  • K, β (GRPO group size and KL coefficient)
    Policy optimization hyperparameters for relative advantages and reference-policy regularization.
  • joint cap 96 and sequence padding policy
    Engineering choices that define the batched representation and can affect topology coverage.
  • skeleton extraction success threshold (reconstruction error cutoff)
    Defines the reported ~85% success rate and which sequences enter training.
axioms (5)
  • domain assumption Articulated mesh motion in the training data is well explained by a fixed skeleton topology and fixed skinning weights with time-varying rigid joint transforms under linear blend skinning.
    Stated in Stage-1 overview and used throughout fitting and inference; fails for highly non-rigid or topology-changing motion (also noted in limitations).
  • domain assumption Automatic first-frame rigging (UniRig) produces a usable joint hierarchy and skinning weights that can be frozen across the sequence.
    Eq. (1) and inference pipeline; quality of all pseudo-labels is bounded by this tool.
  • ad hoc to paper Vertex-wise MSE between LBS-reconstructed and ground-truth meshes is a sufficient objective for recovering motion-relevant joint poses.
    Eqs. (3)–(4); alternatives (annealing, least squares) are ablated but the objective itself is a modeling choice, not a theorem.
  • domain assumption CLIP text embeddings plus autoregressive latent pose prediction capture the intended motion semantics of the engineered prompt templates.
    Stage-2 text conditioning and dataset prompt design sections.
  • standard math Standard transformer residual blocks, masked pooling, and GRPO relative advantages are valid optimization machinery for sequence generation.
    Used as off-the-shelf learning tools without new theoretical claims.
invented entities (2)
  • Motion-GRPO reward suite (smoothness, dynamics, bone consistency) no independent evidence
    purpose: Provide sequence-level preference signals beyond frame-wise reconstruction for temporally coherent articulated motion.
    Paper-specific reward design inspired by GRPO in LLMs; independent evidence is only the ablation table, not external theory.
  • Temporally consistent pseudo-skeleton trajectories from differentiable LBS fitting no independent evidence
    purpose: Replace missing per-frame skeleton annotations with weakly supervised targets for Stage-2 training.
    Core methodological construct; validated by reconstruction MSE and qualitative skeleton comparisons, not by external ground-truth joints on Diffusion4D.

pith-pipeline@v1.1.0-grok45 · 20829 in / 3603 out tokens · 35373 ms · 2026-07-10T10:41:36.413342+00:00 · methodology

0 comments
read the original abstract

We study 4D generation to synthesize temporally coherent sequences of 3D geometry for animation and content creation. In contrast to existing SDS-based optimization methods and video-driven animation approaches, we adopt a skeleton-driven animation framework aligned with standard industrial pipelines, which enables explicit control and editing. To this end, we propose SkelGen4D, a weakly supervised feed-forward framework for text-driven mesh animation that generates explicit skeleton motions without requiring per-frame skeleton annotations. SkelGen4D first recovers temporally consistent pseudo-skeletons from animated meshes via differentiable fitting, and then generates text-conditioned skeleton motion sequences in a feed-forward manner, further refined with Motion-GRPO to ensure temporally coherent, physically plausible, and articulated animation. We evaluate our method on two large-scale benchmarks, Truebones Zoo and Diffusion4D. Our results show that our weakly supervised skeleton modeling matches or surpasses fully supervised baselines while scaling to diverse object categories for high-quality text-driven mesh animation. Further, our method supports flexible motion editing and is aligned with standard animation production pipelines.

Figures

Figures reproduced from arXiv: 2607.08246 by Bin Sheng, Dian Zhang, Hao Feng, Haoran Xie, Jia-Hui Pan, Jingyu Hu, Ka-Hei Hui, Zhengzhe Liu, Zhi Zuo.

Figure 1
Figure 1. Figure 1: Text-driven mesh animation results generated by [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SkelGen4D. Our framework consists of two stages. Stage 1: Weakly supervised skeleton extraction. Given the first-frame mesh, we apply automatic rigging to obtain the skeleton topology and skinning weights W. Then we optimize per-frame skeleton Pi via gradient-based fitting so that the mesh reconstructed by linear blend skinning Mˆ matches the ground truth mesh M sequence, yielding temporally co… view at source ↗
Figure 3
Figure 3. Figure 3: Inference pipeline of SkelGen4D. Given a text prompt and a first-frame mesh, we first obtain an initial skeleton and skinning weights via automatic rigging (UniRig). The text and skeleton are then encoded and fed into an autoregressive transformer to generate a sequence of latent skeleton poses, which are decoded into explicit skeleton motions. Finally, the generated skeleton sequence is applied to the mes… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparisons of skeleton motion generation on the Truebones Zoo dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparisons of 4D generation on the Diffusion4D dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparisons of the annotated skeleton and the predicted pseudo skeleton using our SkelGen4D on the Diffusion4D [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of the controllability of our skeleton-based 4D generative framework SkelGen4D. (a,b) Motion composition. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative ablation results comparing our SGD-based pseudo-skeleton generation with alternative strategies. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗

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

44 extracted references · 44 canonical work pages · 4 internal anchors

  1. [1]

    Fbx skeletal mesh pipeline,

    E. G. D. Documentation, “Fbx skeletal mesh pipeline,” https://dev.epicgames.com/documentation/en-us/unreal-engine/ fbx-skeletal-mesh-pipeline-in-unreal-engine, unreal Engine official documentation

    work page

  2. [2]

    Skeletal animation: A comprehensive guide,

    GarageFarm.net, “Skeletal animation: A comprehensive guide,” https: //garagefarm.net/blog/skeletal-animation-a-comprehensive-guide, indus- trial overview of skeletal animation techniques

    work page

  3. [3]

    DreamGaussian4D: Generative 4D Gaussian Splatting

    J. Ren, L. Pan, J. Tang, C. Zhang, A. Cao, G. Zeng, and Z. Liu, “Dreamgaussian4d: Generative 4d gaussian splatting,”arXiv preprint arXiv:2312.17142, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

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

    H. Ling, S. W. Kim, A. Torralba, S. Fidler, and K. Kreis, “Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    work page 2024

  5. [5]

    V2m4: 4d mesh animation reconstruction from a single monocular video,

    J. Chen, B. Zhang, X. Tang, and P. Wonka, “V2m4: 4d mesh animation reconstruction from a single monocular video,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025

    work page 2025

  6. [6]

    Animateanymesh: A feed-forward 4d foundation model for text-driven universal mesh animation,

    Z. Wu, C. Yu, F. Wang, and X. Bai, “Animateanymesh: A feed-forward 4d foundation model for text-driven universal mesh animation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025

    work page 2025

  7. [7]

    Puppeteer: Rig and animate your 3d models,

    C. Song, X. Li, F. Yang, Z. Xu, J. Wei, F. Liu, J. Feng, G. Lin, and J. Zhang, “Puppeteer: Rig and animate your 3d models,” inAdvances in Neural Information Processing Systems (NeurIPS), 2025

    work page 2025

  8. [8]

    Animax: Animating the inanimate in 3d with joint video-pose diffusion models,

    Z. Huang, H. Feng, Y . Sun, Y . Guo, Y . Cao, and L. Sheng, “Animax: Animating the inanimate in 3d with joint video-pose diffusion models,” inACM SIGGRAPH Asia 2025 Conference Proceedings, 2025

    work page 2025

  9. [9]

    MoCapAnything: Unified 3D Motion Capture for Arbitrary Skeletons from Monocular Videos

    K. Gong, Z. Wen, W. He, M. Xu, Q. Wang, N. Zhang, Z. Li, D. Lian, W. Zhao, X. He, and M. Zhang, “Mocapanything: Unified 3d motion capture for arbitrary skeletons from monocular videos,”arXiv preprint arXiv:2512.10881, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Articulated kinematics distillation from video diffusion models,

    X. Li, Q. Ma, T.-Y . Lin, Y . Chen, C. Jiang, M.-Y . Liu, and D. Xiang, “Articulated kinematics distillation from video diffusion models,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    work page 2025

  11. [11]

    Anytop: Character animation diffusion with any topology,

    I. Gat, S. Raab, G. Tevet, A. H. Bermano, and D. Cohen-Or, “Anytop: Character animation diffusion with any topology,” inACM SIGGRAPH 2025 Conference Proceedings, 2025

    work page 2025

  12. [12]

    One model to rig them all: Diverse skeleton rigging with unirig,

    J.-P. Zhang, C.-F. Pu, M.-H. Guo, Y .-P. Cao, and S.-M. Hu, “One model to rig them all: Diverse skeleton rigging with unirig,”ACM Transactions on Graphics (TOG), 2025

    work page 2025

  13. [13]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . Li, Y . Wuet al., “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,”arXiv preprint arXiv:2402.03300, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  14. [14]

    4d-fy: Text-to-4d generation using hybrid score distillation sampling,

    S. Bahmani, I. Skorokhodov, V . Rong, G. Wetzstein, L. Guibas, P. Wonka, S. Tulyakov, J. J. Park, A. Tagliasacchi, and D. B. Lindell, “4d-fy: Text-to-4d generation using hybrid score distillation sampling,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    work page 2024

  15. [15]

    Consistent4d: Consistent 360◦ dynamic object generation from monocular video,

    Y . Jiang, L. Zhang, J. Gao, W. Hu, and Y . Yao, “Consistent4d: Consistent 360◦ dynamic object generation from monocular video,” inInternational Conference on Learning Representations (ICLR), 2024

    work page 2024

  16. [16]

    Stag4d: Spatial-temporal anchored generative 4d gaussians,

    Y . Zeng, Y . Jiang, S. Zhu, Y . Lu, Y . Lin, H. Zhu, W. Hu, X. Cao, and Y . Yao, “Stag4d: Spatial-temporal anchored generative 4d gaussians,” in European Conference on Computer Vision (ECCV), 2024

    work page 2024

  17. [17]

    Eg4d: Explicit generation of 4d object without score distillation,

    Q. Sun, Z. Guo, Z. Wan, J. N. Yan, S. Yin, W. Zhou, J. Liao, and H. Li, “Eg4d: Explicit generation of 4d object without score distillation,” in International Conference on Learning Representations (ICLR), 2025

    work page 2025

  18. [18]

    Diffusion 2: Dynamic 3d content generation via score composition of video and multi-view diffusion mod- els,

    Z. Yang, Z. Pan, C. Gu, and L. Zhang, “Diffusion 2: Dynamic 3d content generation via score composition of video and multi-view diffusion mod- els,” inInternational Conference on Learning Representations (ICLR), 2025

    work page 2025

  19. [19]

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

    Y . Xie, C.-H. Yao, V . V oleti, H. Jiang, and V . Jampani, “Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency,” in International Conference on Learning Representations (ICLR), 2025

    work page 2025

  20. [20]

    Sv4d 2.0: Enhancing spatio-temporal consistency in multi-view video diffusion for high-quality 4d generation,

    C.-H. Yao, Y . Xie, V . V oleti, H. Jiang, and V . Jampani, “Sv4d 2.0: Enhancing spatio-temporal consistency in multi-view video diffusion for high-quality 4d generation,” inProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), 2025

    work page 2025

  21. [21]

    Cat4d: Create anything in 4d with multi-view video diffusion models,

    R. Wu, R. Gao, B. Poole, A. Trevithick, C. Zheng, J. T. Barron, and A. Holynski, “Cat4d: Create anything in 4d with multi-view video diffusion models,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    work page 2025

  22. [22]

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

    H. Zhang, X. Chen, Y . Wang, X. Liu, Y . Wang, and Y . Qiao, “4diffusion: Multi-view video diffusion model for 4d generation,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

    work page 2024

  23. [23]

    L4gm: Large 4d gaussian reconstruction model,

    J. Ren, K. Xie, A. Mirzaei, H. Liang, X. Zeng, K. Kreis, Z. Liu, A. Torralba, S. Fidler, and S. W. Kim, “L4gm: Large 4d gaussian reconstruction model,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

    work page 2024

  24. [24]

    Animate3d: Animating any 3d model with multi-view video diffusion,

    Y . Jiang, C. Yu, C. Cao, F. Wang, W. Hu, and J. Gao, “Animate3d: Animating any 3d model with multi-view video diffusion,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

    work page 2024

  25. [25]

    Motiondreamer: Exploring semantic video diffusion features for zero-shot 3d mesh animation,

    L. Uzolas, E. Eisemann, and P. Kellnhofer, “Motiondreamer: Exploring semantic video diffusion features for zero-shot 3d mesh animation,” in International Conference on 3D Vision (3DV), 2025

    work page 2025

  26. [26]

    Ss4d: Native 4d generative model via structured spacetime latents,

    Z. Li, M. Zhang, T. Wu, J. Tan, J. Wang, and D. Lin, “Ss4d: Native 4d generative model via structured spacetime latents,”ACM Transactions on Graphics (TOG), 2025

    work page 2025

  27. [27]

    Actionmesh: Animated 3d mesh generation with temporal 3d diffusion,

    R. Sabathier, D. Novotny, N. Mitra, and T. Monnier, “Actionmesh: Animated 3d mesh generation with temporal 3d diffusion,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026

    work page 2026

  28. [28]

    Swit-4d: Sliding-window transformer for lossless and parameter-free temporal 4d generation,

    K. Gong, Z. Wen, M. Xu, W. He, Q. Wang, N. Zhang, Z. Li, C. Li, D. Lian, W. Zhao, X. He, and M. Zhang, “Swit-4d: Sliding-window transformer for lossless and parameter-free temporal 4d generation,”

    work page

  29. [29]

    Available: https://arxiv.org/abs/2512.10860

    [Online]. Available: https://arxiv.org/abs/2512.10860

    work page arXiv

  30. [30]

    Motionclip: Exposing human motion generation to clip space,

    G. Tevet, B. Gordon, A. Hertz, A. H. Bermano, and D. Cohen-Or, “Motionclip: Exposing human motion generation to clip space,” in European Conference on Computer Vision (ECCV), 2022

    work page 2022

  31. [31]

    Human motion diffusion model,

    G. Tevet, S. Raab, B. Gordon, Y . Shafir, D. Cohen-Or, and A. H. Bermano, “Human motion diffusion model,” inInternational Conference on Learning Representations (ICLR), 2023

    work page 2023

  32. [32]

    Motiondiffuse: Text-driven human motion generation with diffusion model,

    M. Zhang, Z. Cai, L. Pan, F. Hong, X. Guo, L. Yang, and Z. Liu, “Motiondiffuse: Text-driven human motion generation with diffusion model,”IEEE Transactions on Pattern Analysis and Machine Intelli- gence (TPAMI), 2024

    work page 2024

  33. [33]

    Motiongpt: Hu- man motion as a foreign language,

    B. Jiang, X. Chen, W. Liu, J. Yu, G. Yu, and T. Chen, “Motiongpt: Hu- man motion as a foreign language,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

    work page 2023

  34. [34]

    Large motion model for unified multi- modal motion generation,

    M. Zhang, D. Jin, C. Gu, F. Hong, Z. Cai, J. Huang, C. Zhang, X. Guo, L. Yang, Y . He, and Z. Liu, “Large motion model for unified multi- modal motion generation,” inEuropean Conference on Computer Vision (ECCV), 2024

    work page 2024

  35. [35]

    Smpl: A skinned multi-person linear model,

    M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black, “Smpl: A skinned multi-person linear model,”ACM Transactions on Graphics (TOG), 2015

    work page 2015

  36. [36]

    Skele- ton extraction by mesh contraction,

    O. K.-C. Au, C.-L. Tai, H.-K. Chu, D. Cohen-Or, and T.-Y . Lee, “Skele- ton extraction by mesh contraction,”ACM Transactions on Graphics (TOG)), 2008

    work page 2008

  37. [37]

    l1-medial skeleton of point cloud,

    H. Huang, S. Wu, D. Cohen-Or, M. Gong, H. Zhang, G. Li, and B. Chen, “l1-medial skeleton of point cloud,”ACM Transactions on Graphics (TOG), 2013

    work page 2013

  38. [38]

    Morig: Motion-aware rigging of character meshes from point clouds,

    Z. Xu, Y . Zhou, L. Yi, and E. Kalogerakis, “Morig: Motion-aware rigging of character meshes from point clouds,” inSIGGRAPH Asia, 2022

    work page 2022

  39. [39]

    Magicarticulate: Make your 3d models articulation-ready,

    C. Song, J. Zhang, X. Liet al., “Magicarticulate: Make your 3d models articulation-ready,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    work page 2025

  40. [40]

    Riganything: Template-free autoregressive rigging for diverse 3d assets,

    I. Liu, Z. Xu, Y . Wang, H. Tan, Z. Xu, X. Wang, H. Su, and Z. Shi, “Riganything: Template-free autoregressive rigging for diverse 3d assets,”ACM Transactions on Graphics (TOG), 2025

    work page 2025

  41. [41]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning (ICML), 2021

    work page 2021

  42. [42]

    Truebones fbx/.bvh zoo (over 75 animals with animations and textures),

    Truebones Motions Animation Studios, “Truebones fbx/.bvh zoo (over 75 animals with animations and textures),” https://truebones.gumroad. com/, gumroad product page

    work page

  43. [43]

    Diffusion4D: Fast Spatial-temporal Consistent 4D Generation via Video Diffusion Models

    H. Liang, Y . Yin, D. Xu, H. Liang, Z. Wang, K. N. Plataniotis, Y . Zhao, and Y . Wei, “Diffusion4d: Fast spatial-temporal consistent 4d generation via video diffusion models,”arXiv preprint arXiv:2405.16645, 2024. 12

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  44. [44]

    Single motion diffusion,

    S. Raab, I. Leibovitch, G. Tevet, M. Arar, A. H. Bermano, and D. Cohen- Or, “Single motion diffusion,” inInternational Conference on Learning Representations (ICLR), 2024. VI. BIOGRAPHYSECTION Hao FengHao Feng received the B.S. degree from the Nanjing University of Information Science and Technology (NUIST), Nanjing, China, in 2022. He has entered the sch...

    work page 2024