Pith. sign in

REVIEW 4 major objections 4 minor 49 references

Scheduled inpainting blends a source animation into every denoising step, letting pretrained motion diffusion models edit existing clips interactively for the first time.

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

2026-08-03 13:02 UTC pith:YCJ3J5X5

load-bearing objection A straightforward inference-time method for editing existing motion clips with pretrained diffusion models; the core idea is real and practically useful, but the quantitative evaluation is thinner and the universality claim softer than the prose suggests. the 4 major comments →

arxiv 2607.29133 v1 pith:YCJ3J5X5 submitted 2026-07-31 cs.GR

Interactive Generative Motion Editing via Scheduled Inpainting

Dhruv Agrawal , Dominik Borer , Luca V\"ogeli , Robert Sumner , Martin Guay , Jakob Buhmann This is my paper
classification cs.GR
keywords scheduled inpaintingmotion editingdiffusion modelsinteractive animationmotion inpaintinggenerative motiondirect manipulationmotion stitching
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.

This paper tries to establish that a pretrained generative motion model — one trained to create new animation from scratch — can instead be used to edit an existing animation clip interactively, preserving its style while allowing large structural changes. The proposed inference-time technique, scheduled inpainting, works by blending the source clip into the model's own denoising output at every diffusion step, with a user-controlled schedule and a per-joint, per-frame mask deciding how much of the original motion to keep. Because the blending happens in a normalized alignment space and costs nothing beyond the usual generation pass, the same framework can extend, stitch, composite, retime, and directly drag points on real clips without retraining the model. To the authors' knowledge this is the first interactive system for generative motion editing, and they validate it against four baselines, ablations, and professional artist feedback.

Core claim

The paper's central claim is that scheduling the amount of inpainting over the diffusion process — rather than using a fixed or binary mask — unlocks interactive generative motion editing. At high noise levels the base motion is copied wholesale into the generation, at low noise levels the model is left alone, and in between a linear interpolation controlled by two user parameters σ_s and σ_e blends the two. The same formula, with a spatiotemporal mask α_mask, decides which joints and frames are preserved. The enabling insight is that this blend must be computed in a custom alignment space — origin at the first frame, facing along the x-axis, zero mean and unit variance, differential root co

What carries the argument

The load-bearing object is the scheduled blend identity M0_gen = α_t M_base + (1−α_t) M0_gen, evaluated at every denoising step, where α_t = α_time^t · α_mask. The timestep schedule α_time^t is a linear ramp between user-set noise levels σ_s and σ_e (full overwrite above σ_s, no inpainting below σ_e), giving direct control over preservation versus generation. The spatiotemporal mask α_mask, built from Gaussian kernels around user constraints, localizes editing to chosen joints and frames. All blending happens in a hand-designed inpainting space that aligns both sequences to a common origin and orientation, normalizes scale, and represents the root via differential coordinates — this normaliz

Load-bearing premise

The method assumes that, after normalizing scale, orientation, and root drift, linearly interpolating a source clip into the model's generated estimate at every denoising step stays on the manifold of motions the pretrained model can still denoise naturally.

What would settle it

Stitch a crouching clip with a long-stride clip of the same character via scheduled inpainting and inspect the blended poses at high-noise denoising steps: if foot sliding, self-penetration, or limb spasm appears despite the alignment and normalization, the linear-blend assumption fails. A direct quantitative check is to compare root-relative reconstruction error with and without the paper's normalization on a scale-mismatched pair — the paper predicts normalization removes most of the error.

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

If this is right

  • Any pretrained interactive motion diffusion model can be turned into an editing tool with no retraining and no added inference cost beyond the standard generation pass.
  • Artists can chain and combine editing operations — extending a clip, stitching two clips, compositing upper and lower body, retiming, and dragging points — inside one real-time session while the base motion stays intact.
  • Because the source clip is inpainted rather than passed as a hard constraint, the edited result itself remains editable, enabling iterative refinement.
  • The preservation/generation trade-off is explicitly user-controllable through two scalars, and reconstruction error improves monotonically as inpainting strength increases.
  • Scheduled inpainting succeeds in cases where noise-inversion and noise-optimization approaches fail or run too slowly for interactive use (the paper reports roughly 0.19 seconds per batch versus 4.5–10.3 seconds for inversion baselines).

Where Pith is reading between the lines

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

  • The same schedule-and-blend recipe should transfer to other generative modalities — audio, video, shape — wherever a pretrained diffusion prior exists and the task is 'edit an existing exemplar without retraining.'
  • The shrinking Gaussian kernel for sparse constraints suggests a general principle worth extracting: constraint influence should be annealed with noise level so that high noise steers global structure and low noise enforces local detail.
  • The paper's own convex-hull observation implies a practical ceiling: edits are confined to the model's learned distribution, so a feasibility visualizer (which the authors mention as future work) may become a necessary companion in production use.
  • A testable extension: apply the same normalization and scheduled blend to stylized or non-human characters; the claimed model-agnosticism predicts it should work with any diffusion backbone that accepts joint-and-frame conditioning.

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

4 major / 4 minor

Summary. The paper introduces scheduled inpainting, an inference-time method for editing existing motion clips with pretrained interactive motion diffusion models. The core idea is to blend the base motion M_base into the denoised generation at every diffusion step via Eq. (1), with a user-controllable schedule α_time and a spatiotemporal preservation mask α_mask. A dedicated inpainting space (Section 3.3) aligns and normalizes both sequences before blending. The method is demonstrated on direct manipulation, extension, stitching, compositing, and retiming, using IBMM and SF-control as backbone models. The evaluation comprises qualitative comparisons against MotionLab, CondEditor, DNO, and noise-inversion, quantitative reconstruction metrics in Tables 1 and 2, ablations of the schedule, mask, normalization, and kernel shape, and a usability test with two professional artists.

Significance. If the results hold, the paper makes a useful practical contribution: a training-free, inference-only technique that upgrades pretrained generative motion models into interactive editing tools, potentially reducing manual effort in animation production. The method is simple, conceptually clear, and the authors are honest about limitations, including the 'convex hull problem' in Appendix C. The strengths are the compatibility with multiple backbone models, the unified treatment of several editing tasks, and the inclusion of ablations. However, the quantitative evaluation is thin and partly circular: the main reconstruction metrics measure distance to the base motion that Eq. (1) explicitly injects, and the user study has only two participants. These weaknesses currently make the broad claims of 'extensive validation' and a 'single interactive framework' stronger than the evidence supports.

major comments (4)
  1. [§5.3.1, Table 2, Eq. (1)] The L2P/L2R reconstruction metrics measure distance to M_base, which Eq. (1) directly injects at high noise through α_t. The ablation across σ_s and σ_e is therefore partly tautological: stronger inpainting schedules mechanically lower the error to the base motion. The statement that σ_s=500/σ_e=50 'best balances' preservation and natural edits is not supported by these numbers alone; it relies on qualitative inspection. Please add evaluation metrics that separate preservation from edit quality (e.g., constraint satisfaction, physical plausibility, user preference) and report variances across random seeds and motion clips.
  2. [§5.4] The usability test involves only two professional artists and is entirely qualitative. This is too thin to support the abstract's 'extensively validate our approach' or the broad claim of enabling a production-ready workflow. A larger user study with statistical analysis, or at minimum a clearly labeled pilot study with softened conclusions, is needed. As it stands, the evidence from the usability test is anecdotal.
  3. [Appendix C, §6] Appendix C shows that for exaggerated or periodic motions, the plain scheduled inpainting of Eq. (1) loses the base motion's dynamics; the authors must automatically detect and add extrema points as additional constraints. This directly demonstrates that the inpainting space of Section 3.3 does not in general keep the blended sequence on the model's motion manifold. The paper should either integrate this constraint-addition into the method description and evaluate it as part of the framework, or explicitly scope the method's applicability and characterize failure modes. Otherwise the claim of a 'single interactive framework' for arbitrary input clips is overbroad.
  4. [§5.2.2–5.2.3] The comparisons with MotionLab, CondEditor, DNO, and noise-inversion are mostly qualitative or based on a single clip without error bars. Since the central novelty is a general inference-time framework, the absence of a quantitative comparison over a diverse set of clips, with distributions over seeds and clips, limits the support for the claimed advantage over these baselines. Please provide quantitative results with variance for at least the interactive baselines, or temper the comparative claims accordingly.
minor comments (4)
  1. [Eq. (4)–(5)] Notation inconsistency: G(j,t;c_i) is defined in Eq. (4) but the expression depends only on t and c_i, and Eq. (5) uses G(t;c_i). Align the notation to avoid confusion.
  2. [§5.3.4, Figure 13] The text refers to a 'piecewise linear' kernel while Figure 13 labels the same kernel as 'Triangular kernel'. Use one consistent name.
  3. [Table 3] The 'Ours' row lists σ_s=500 and σ_e=50, but the recommended schedule is discussed in Table 2. Ensure the reader can easily map the schedule notation between tables.
  4. [Abstract and §5] The abstract states 'extensively validate our approach', but the quantitative evaluation lacks error bars and the user study has two participants. Consider softening the wording to match the evidence presented.

Circularity Check

1 steps flagged

Reconstruction validation is tautological: L2P/L2R measure distance to M_base, which Eq. (1) injects directly; otherwise the pipeline is self-contained.

specific steps
  1. self definitional [Section 5.3.1, Table 2 (and Table 1; Eq. 1, Section 3)]
    "To evaluate the results, we measure foot-sliding and root-relative reconstruction error of non-root joints, reported as L2 position error (L2P) and L2 rotation error (L2R), to quantify the motion naturalness and similarity to base motion. ... For stronger inpainting schedules, the reconstruction error improves linearly with the strongest schedule of σs=300 and σe=50 achieving the best metrics."

    Eq. (1) defines the inpainted motion as a convex combination α_t×M_base+(1−α_t)×M0_gen. L2P/L2R are exactly the distance to M_base. Therefore increasing α (stronger inpainting schedule) mechanically decreases L2P/L2R; the monotone improvement reported in Table 2 is the injected fraction, not an emergent property of the generative prior or the representation space. The same metric supports the 'significantly improving base motion preservation' statement in Table 1. The ablation is thus a measure of the user-controlled injection knob, not independent evidence that the method reconstructs or preserves motion. DNO/noise-inversion comparisons on this metric are similarly tilted, because scheduled inpainting is allowed to copy M_base into the output by construction.

full rationale

Scheduled inpainting itself is mostly self-contained: Eqs. (1)-(5) define an inference-time blend and spatiotemporal mask; Section 3.3's alignment/normalization is explicitly designed and ablated (Figs. 11, 16); and the comparisons to DNO, noise-inversion, MotionLab, and CondEditor, the foot-sliding metric, and the artist usability test provide independent evidence for the interactive-editing claim. The self-citations to [ABB*24] (SKEL-Betweener) and [VAG*25] (IBMM) are used as a prior rig/model and to construct a baseline, not to prove the central mechanism, so they are not load-bearing circularity. Appendix C is an honest limitation: highly exaggerated/periodic motions fall outside the model's distribution and require automatic extrema constraints, which weakens the universal-interactive-editing claim but is not a circular step. The one genuine circularity is the quantitative reconstruction validation: because the output is a weighted copy of M_base, L2P/L2R vs M_base are controlled by the schedule, making the Table 2 ablation and part of Table 1 tautological. This does not invalidate the whole system—the qualitative, foot-sliding, and user-study evidence stands—but the headline preservation metrics reduce by construction. Score 5.0.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The method rests on pretrained interactive motion diffusion models and on the assumption that blending in a zero-mean/unit-variance, root-aligned space stays within the manifold of plausible motions. The schedule endpoints and influence radius are user-tunable and selected by ablation, constituting free parameters. No new physical or conceptual entities are introduced.

free parameters (3)
  • σ_s (schedule start) = 500 (recommended)
    Diffusion timestep above which the base motion completely overwrites the generated motion. Chosen via ablation in §5.3.1 to balance preservation and naturalness on the internal and LaFan1 datasets.
  • σ_e (schedule end) = 50 (recommended)
    Diffusion timestep below which no inpainting is applied. Chosen in the same ablation; the paper notes that some inference steps without inpainting are needed to resolve interpolation artifacts.
  • µ_ci (influence radius) = 5–40 (user-controlled)
    Width of the Gaussian kernel in Eq. (4) that controls the temporal extent of a constraint's influence. The paper shows the effect for µ=5,10,20,40 in §5.3.3 but does not provide an automated selection rule.
axioms (3)
  • domain assumption Pretrained models IBMM [VAG*25] and SF-control [HBLK25] exist, run at interactive rates, and expose the model's predicted clean motion M0_gen at each diffusion timestep.
    The entire method depends on these external models' availability and interface. They are cited but not provided; the paper does not fully specify how M0_gen is extracted from each model.
  • domain assumption Linear interpolation in the paper's normalized inpainting space yields inputs that the diffusion model can still denoise plausibly.
    Section 3.3 assumes that zero-mean, unit-variance scaling and root alignment are sufficient to avoid artifacts when mixing clips with different global orientations or scales; this is an empirical assumption not proven against the model's training distribution.
  • domain assumption The 'convex hull problem' described in Appendix C is a real limitation of the underlying generative models.
    The automatic addition of extrema constraints as a remedy relies on this empirical characterization of model behavior. The paper provides qualitative evidence but no formal analysis.

pith-pipeline@v1.3.0-daily-deepseek · 15513 in / 9846 out tokens · 103340 ms · 2026-08-03T13:02:32.255941+00:00 · methodology

0 comments
read the original abstract

Motion editing is central to VFX and game development, where it is used extensively to modify and augment existing movements to conform to new environments or changes in artistic direction. While traditional motion editing can do small modifications, it cannot accommodate larger structural edits, resulting in visual warping artifacts that require authoring new motion. Conversely, recent advances in large-scale generative modeling have unlocked newfound capabilities for authoring entire movements by directly manipulating sparse spatial constraints. While impressive at creating new movements, these methods lack the capability to preserve and edit existing motion interactively. In this work, we introduce scheduled inpainting, a method that enables interactive generative motion editing, a novel paradigm unifying motion synthesis and editing by leveraging generative models. Scheduled inpainting is a simple yet powerful inference-based technique that enables fine-grained spatiotemporal control over the balance between preserving the original motion and generating new content. By building atop generative models that support direct manipulation, our system allows artists to interactively refine existing animations while ensuring results remain natural and consistent with the learned motion distribution. Scheduled inpainting is versatile and supports many editing applications, such as extending, stitching, and compositing different clips. Finally, we extensively validate our approach by comparing with four baselines, conducting ablations of our design, and reporting user feedback.

Figures

Figures reproduced from arXiv: 2607.29133 by Dhruv Agrawal, Dominik Borer, Jakob Buhmann, Luca V\"ogeli, Martin Guay, Robert Sumner.

Figure 1
Figure 1. Figure 1: Scheduled inpainting allows for multiple generative motion editing tasks within the same real-time framework, while resulting in natural motion. We see here two clips in blue and red being stitched together (purple), extended (grey), as well as directly modified by dragging points (yellow dots). Abstract Motion editing is central to VFX and game development, where it is used extensively to modify and augme… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our scheduled inpainting for interactive generative motion editing. A user can specify which motion should be used (red and blue), as well as control where and how much the motion should be preserved via controls impacting the spatiotemporal mask, which in turn dictates the schedule. training a specific model, in contrast to our approach that works at inference with different pretrained diffusi… view at source ↗
Figure 4
Figure 4. Figure 4: An example of the spatiotemporal mask αmask for a single joint when a user manipulates two constraints at frames 35 and 55. The motion is allowed to be generated in these areas, and is steered towards preservation in other parts of the timeline. but walk away in opposite directions results in an average motion that is mostly static near the starting position. Or mixing two clips that are of very different … view at source ↗
Figure 5
Figure 5. Figure 5: Generative motion editing allows sampling different mo￾tions, such as different extensions of a rolling motion. Each curve represents the motion path for an end-effector joint and the two black silhouettes represent the start and end of the base motion. Note that Mbase must be padded with additional frames to align with αextend in Eq. (1). In [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Motion editing examples using point-based constraints (green) for AMASS (left), LaFan1 (middle) and internal (right) datasets. Original motion is shown in cyan and edited motion in yellow. only warps the motion and cannot accommodate structural changes when timings change significantly. By setting αmask = 0.8 · 1J×T , our scheduled inpainting method uses the warped motion as the base motion to recover natu… view at source ↗
Figure 8
Figure 8. Figure 8: With scheduled inpainting, we can directly manipulate a few joints to modify a jumping motion (left) to include a natural kick in midair (right). The manipulated constraints are highlighted in green and also shown in our video. 5.2.1. Results First, we use the IBMM [VAG∗ 25] motion model to edit a jumping motion to include a kick in midair, shown in [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: We edit the end frame (shown in red) of a rolling forward clip to go behind the starting frame (shown in black). Each row shows the output from noise-inversion [SME22] (top), DNO [KPA∗ 24] (middle), and our scheduled inpainting method (bottom) at three different moments or frames of the motion. While noise-inversion and DNO give an unnatural turn that is too fast in the first two frames, our method generat… view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of scheduled inpainting (right) to Motion￾Lab [GHSZ25] (left). MotionLab cannot extend the walking dis￾tance (a) and ignores sparse constraints (green) (b), while our ap￾proach follows them accurately. using sparse constraints, instead of dense trajectories, as shown in green in Figure 10b, where we edit the arm position. MotionLab completely ignores the constraints while the SF-control model u… view at source ↗
Figure 11
Figure 11. Figure 11: Effect of aligning motions during inpainting. Without alignment, the character unnaturally moves towards the original keyframe (in cyan) before warping to the target keyframe. In con￾trast, with alignment the motion remains natural and consistent. Lastly, CondEditor similarly exhibits large reconstruction errors while struggling to perform realistic edits, as described earlier. 5.3.2. Motion Alignment In … view at source ↗
Figure 12
Figure 12. Figure 12: We can see the effect of the influence radius µci when editing to avoid an obstacle. Notice how more frames are adapted to avoid the obstacle, as the radius increases. (a) Base motion (b) Square kernel (c) Triangular kernel (d) Gaussian kernel [PITH_FULL_IMAGE:figures/full_fig_p009_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: Before (left) and after (right) fixing scene penetration and stitching artifacts in the artist test scene. 5.4. Usability Test To evaluate the usability of our method, we conducted a test with two professional digital artists tasked with animating a character through a parkour course. This task required running, crawling, and jumping, using a library of two running clips and one crawling mo￾tion clip. As … view at source ↗
Figure 13
Figure 13. Figure 13: Ablating the weight decay kernel function for direct ma￾nipulation. While the piecewise linear and gaussian kernels can recreate the roll, the square wave results in a ducking motion. require that the partial inpainting at the kernel’s edges sufficiently anchors the generated motion to the base motion, so the edit follows a trajectory close to the original rather than diverging arbitrarily. Additionally, … view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of using sparse constraints with Cond￾MDI [CTR∗ 24]. By default, CondMDI completely ignores the sparse hip constraint (in green), while using standard inpainting results in large spikes. Using scheduled inpainting better follows the constraint while creating smooth motion. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_15.png] view at source ↗
Figure 17
Figure 17. Figure 17: From an exaggerated base motion (left), we can see how automatically adding some sparse constraints with our method (shown on the right), better re-constructs the motion than without any constraint (shown in the middle). closely those of the cyan character than those of the middle char￾acter. While those additional constraints help to better match the original base motion, they also come at the cost of ma… view at source ↗
Figure 16
Figure 16. Figure 16: We edit a running clip to run half the distance of the original sequence. Without normalization, the base motion results in an unnatural warp to the final frame with foot sliding. In con￾trast, with normalization, the resulting motion does not have such artifacts. Appendix B: Ablating Normalization When moving the boundary frames to stretch or squash an original animation, the overall scale of the animati… view at source ↗
Figure 18
Figure 18. Figure 18: Number of steps required for noise-inversion. Evaluating too few steps results in the inverted motion (white) not matching the ground truth (cyan) in height. At higher number of evaluation steps, the method is not real-time as shown in our video [PITH_FULL_IMAGE:figures/full_fig_p014_18.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

49 extracted references · 6 linked inside Pith

  1. [1]

    W., Guay M

    Agrawal D., Buhmann J., Borer D., Sumner R. W., Guay M. : SKEL-Betweener : A neural motion rig for interactive motion authoring. ACM Transactions on Graphics (TOG) 43, 6 (2024), 1--11

    2024

  2. [2]

    J., Varol G

    Athanasiou N., Cseke A., Diomataris M., Black M. J., Varol G. : MotionFix : Text-driven 3D human motion editing. In SIGGRAPH Asia 2024 Conference Papers (2024), pp. 1--11

    2024

  3. [3]

    : Autodesk motionbuilder, 2025

    Autodesk Inc. : Autodesk motionbuilder, 2025. Version 2025, Autodesk, Inc., San Rafael, CA. URL: https://www.autodesk.com/products/motionbuilder/

    2025

  4. [4]

    M., Kapadia M

    Bae J., Hwang I., Lee Y.-Y., Guo Z., Liu J., Ben-Shabat Y., Kim Y. M., Kapadia M. : Less is more: Improving motion diffusion models with sparse keyframes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2025), pp. 11069--11078

    2025

  5. [5]

    Chai J., Hodgins J. K. : Constraint-based motion optimization using a statistical dynamic model. In ACM SIGGRAPH 2007 papers. ACM, 2007

    2007

  6. [6]

    : Text-driven human motion generation with motion masked diffusion model

    Chen X. : Text-driven human motion generation with motion masked diffusion model. arXiv preprint arXiv:2409.19686 (2024)

    Pith/arXiv arXiv 2024

  7. [7]

    : Pay attention and move better: Harnessing attention for interactive motion generation and training-free editing

    Chen L.-H., Lu S., Dai W., Dou Z., Ju X., Wang J., Komura T., Zhang L. : Pay attention and move better: Harnessing attention for interactive motion generation and training-free editing. arXiv preprint arXiv:2410.18977 (2024)

    Pith/arXiv arXiv 2024

  8. [8]

    B., van de Panne M

    Cohan S., Tevet G., Reda D., Peng X. B., van de Panne M. : Flexible motion in-betweening with diffusion models. In ACM SIGGRAPH 2024 Conference Papers (2024), pp. 1--9

    2024

  9. [9]

    W., Zhao N

    Guo Z., Hu Z., Soh D. W., Zhao N. : MotionLab : Unified human motion generation and rditing via the motion-condition-motion paradigm. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2025), pp. 13869--13879

    2025

  10. [10]

    : Motion editing with spacetime constraints

    Gleicher M. : Motion editing with spacetime constraints. In Proceedings of the 1997 symposium on Interactive 3D graphics (1997), pp. 139--ff

    1997

  11. [11]

    : Control operators for interactive character animation

    Gou R., van de Panne M., Holden D. : Control operators for interactive character animation. ACM Transactions on Graphics (TOG) 44, 6 (2025), 1--20

    2025

  12. [12]

    K., Fatahalian K

    Goel P., Zhang H., Liu C. K., Fatahalian K. : Generative motion infilling from imprecisely timed keyframes. In Computer Graphics Forum (2025), Wiley Online Library, p. e70060

    2025

  13. [13]

    Hwang I., Bae J., Lim D., Kim Y. M. : Motion synthesis with sparse and flexible keyjoint control. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2025)

    2025

  14. [14]

    URL: https://www.sidefx.com/docs/houdini/character/kinefx/index.html

    Houdini : Kinefx, 2025. URL: https://www.sidefx.com/docs/houdini/character/kinefx/index.html

    2025

  15. [15]

    : Classifier-free diffusion guidance

    Ho J., Salimans T. : Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

    Pith/arXiv arXiv 2022

  16. [16]

    : COMO : Controllable motion generation through language guided pose code editing

    Huang Y., Wan W., Yang Y., Callison-Burch C., Yatskar M., Liu L. : COMO : Controllable motion generation through language guided pose code editing. In European Conference on Computer Vision (2024), Springer, pp. 180--196

    2024

  17. [17]

    G., Yurick M., Nowrouzezahrai D., Pal C

    Harvey F. G., Yurick M., Nowrouzezahrai D., Pal C. : Robust motion in-betweening. ACM Transactions on Graphics (TOG) 39, 4 (2020), 60--1

    2020

  18. [18]

    : Dynamic motion blending for versatile motion editing

    Jiang N., Li H., Yuan Z., He Z., Chen Y., Liu T., Zhu Y., Huang S. : Dynamic motion blending for versatile motion editing. In Proceedings of the Computer Vision and Pattern Recognition Conference (2025), pp. 22735--22745

    2025

  19. [19]

    : Optimizing diffusion noise can serve as universal motion priors

    Karunratanakul K., Preechakul K., Aksan E., Beeler T., Suwajanakorn S., Tang S. : Optimizing diffusion noise can serve as universal motion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024), pp. 1334--1345

    2024

  20. [20]

    : SimMotionEdit : Text-based human motion editing with motion similarity prediction

    Li Z., Cheng K., Ghosh A., Bhattacharya U., Gui L., Bera A. : SimMotionEdit : Text-based human motion editing with motion similarity prediction. In Proceedings of the Computer Vision and Pattern Recognition Conference (2025), pp. 27827--27837

    2025

  21. [21]

    : GENMO : A generalist model for human motion

    Li J., Cao J., Zhang H., Rempe D., Kautz J., Iqbal U., Yuan Y. : GENMO : A generalist model for human motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2025), pp. 11766--11776

    2025

  22. [22]

    : A motion is worth a hybrid sentence: Taming language model for unified motion generation by fine-grained planning

    Li R., Han L., Shu S., Liu Y., Lin Y., Ma Y., Guo J., Liu Z., Li X. : A motion is worth a hybrid sentence: Taming language model for unified motion generation by fine-grained planning. In Proceedings of the 33rd ACM International Conference on Multimedia (2025), pp. 1404--1413

    2025

  23. [23]

    Lee J., Shin S. Y. : A hierarchical approach to interactive motion editing for human-like figures. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques (1999), pp. 39--48

    1999

  24. [24]

    : Drag your noise: Interactive point-based editing via diffusion semantic propagation

    Liu H., Xu C., Yang Y., Zeng L., He S. : Drag your noise: Interactive point-based editing via diffusion semantic propagation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2024), pp. 6743--6752

    2024

  25. [25]

    : Interactive generation of human animation with deformable motion models

    Min J., Chen Y.-L., Chai J. : Interactive generation of human animation with deformable motion models. ACM Transactions on Graphics (TOG) 29, 1 (2009), 1--12

    2009

  26. [26]

    : Contact-aware human motion generation from textual descriptions

    Ma S., Cao Q., Zhang J., Tao D. : Contact-aware human motion generation from textual descriptions. arXiv preprint arXiv:2403.15709 (2024)

    Pith/arXiv arXiv 2024

  27. [27]

    F., Pons-Moll G., Black M

    Mahmood N., Ghorbani N., Troje N. F., Pons-Moll G., Black M. J. : AMASS : Archive of motion capture as surface shapes. In International Conference on Computer Vision (Oct. 2019), pp. 5442--5451

    2019

  28. [28]

    N., Bocquelet F., Harvey F

    Oreshkin B. N., Bocquelet F., Harvey F. G., Raitt B., Laflamme D. : ProtoRes : Proto-residual network for pose authoring via learned inverse kinematics. In International Conference on Learning Representations (2022)

    2022

  29. [29]

    U., Karunratanakul K., Wang P., Xue H., Chen C., Guo C., Cao J., Ren J., Tulyakov S

    Pinyoanuntapong E., Saleem M. U., Karunratanakul K., Wang P., Xue H., Chen C., Guo C., Cao J., Ren J., Tulyakov S. : MaskControl : Spatio-temporal control for masked motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2025)

    2025

  30. [30]

    : High-resolution image synthesis with latent diffusion models

    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), pp. 10684--10695

    2022

  31. [31]

    : MixerMDM : Learnable composition of human motion diffusion models

    Ruiz-Ponce P., Barquero G., Palmero C., Escalera S., Garc \' a-Rodr \' guez J. : MixerMDM : Learnable composition of human motion diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference (2025), pp. 12380--12390

    2025

  32. [32]

    W., Guay M., Buhmann J

    Studer J., Agrawal D., Borer D., Sadat S., Sumner R. W., Guay M., Buhmann J. : Factorized motion diffusion for precise and character-agnostic motion inbetweening. In Proceedings of the 17th ACM SIGGRAPH Conference on Motion, Interaction, and Games (2024), pp. 1--10

    2024

  33. [33]

    Sadat S., Hilliges O., Weber R. M. : Eliminating oversaturation and artifacts of high guidance scales in diffusion models. In The Thirteenth International Conference on Learning Representations (2024)

    2024

  34. [34]

    : Denoising diffusion implicit models

    Song J., Meng C., Ermon S. : Denoising diffusion implicit models. In International Conference on Learning Representations (2022)

    2022

  35. [35]

    : Human motion diffusion as a generative prior

    Shafir Y., Tevet G., Kapon R., Bermano A. : Human motion diffusion as a generative prior. In International Conference on Learning Representations (2024), vol. 2024, pp. 8717--8733

    2024

  36. [36]

    Shi Y., Wang J., Jiang X., Lin B., Dai B., Peng X. B. : Interactive character control with auto-regressive motion diffusion models. ACM Transactions on Graphics (TOG) 43, 4 (2024), 1--14

    2024

  37. [37]

    H., Pan J., Yan H., Zhang W., Tan V

    Shi Y., Xue C., Liew J. H., Pan J., Yan H., Zhang W., Tan V. Y., Bai S. : DragDiffusion : Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024), pp. 8839--8849

    2024

  38. [38]

    : EDGE : Editable dance generation from music

    Tseng J., Castellon R., Liu K. : EDGE : Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023), pp. 448--458

    2023

  39. [39]

    Tessler C., Guo Y., Nabati O., Chechik G., Peng X. B. : MaskedMimic : Unified physics-based character control through masked motion inpainting. ACM Transactions on Graphics (TOG) 43, 6 (2024), 1--21

    2024

  40. [40]

    Tevet G., Raab S., Gordon B., Shafir Y., Cohen-or D., Bermano A. H. : Human motion diffusion model. In The Eleventh International Conference on Learning Representations (2023)

    2023

  41. [41]

    W., Buhmann J

    V \"o geli L., Agrawal D., Guay M., Borer D., Sumner R. W., Buhmann J. : Implicit B \'e zier motion model for precise spatial and temporal control. In Proceedings of the 2025 18th ACM SIGGRAPH Conference on Motion, Interaction, and Games (2025), pp. 1--10

    2025

  42. [42]

    F., Picard D., Kalogeiton V

    Wang X., Dufour N., Andreou N., Cani M.-P., Abrevaya V. F., Picard D., Kalogeiton V. : Analysis of classifier-free guidance weight schedulers. Transactions on Machine Learning Research (2024). URL: https://openreview.net/forum?id=SUMtDJqicd

    2024

  43. [43]

    : Leader is guided: Interactive motion generation via lead-follow paradigm and trajectory guidance

    Wang R., Ma C., Zhao J., Xu H., Sun D., Chen H., Xiong L., Wang Z., Li X. : Leader is guided: Interactive motion generation via lead-follow paradigm and trajectory guidance. In Proceedings of the 33rd ACM International Conference on Multimedia (2025), pp. 1802--1811

    2025

  44. [44]

    : Motion warping

    Witkin A., Popovic Z. : Motion warping. In Proceedings of the 22nd annual conference on Computer graphics and interactive techniques (1995), pp. 105--108

    1995

  45. [45]

    : FineMotion : A dataset and benchmark with both spatial and temporal annotation for fine-grained motion generation and editing

    Wu B., Xie J., Ding M., Kong Z., Ren J., Bai R., Qu R., Shen L. : FineMotion : A dataset and benchmark with both spatial and temporal annotation for fine-grained motion generation and editing. arXiv preprint arXiv:2507.19850 (2025)

    Pith/arXiv arXiv 2025

  46. [46]

    Xu W., Fan S., Henderson P., Ho E. S. : Multi-person interaction generation from two-person motion priors. In Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers (2025), pp. 1--11

    2025

  47. [47]

    : Sketch2Anim : Towards transferring sketch storyboards into 3D animation

    Zhong L., Guo C., Xie Y., Wang J., Li C. : Sketch2Anim : Towards transferring sketch storyboards into 3D animation. ACM Transactions on Graphics (TOG) 44, 4 (2025), 1--15

    2025

  48. [48]

    : FineMoGen : Fine-grained spatio-temporal motion generation and editing

    Zhang M., Li H., Cai Z., Ren J., Yang L., Liu Z. : FineMoGen : Fine-grained spatio-temporal motion generation and editing. Advances in Neural Information Processing Systems 36 (2023), 13981--13992

    2023

  49. [49]

    : Motion anything: Any to motion generation

    Zhang Z., Wang Y., Mao W., Li D., Zhao R., Wu B., Song Z., Zhuang B., Reid I., Hartley R. : Motion anything: Any to motion generation. arXiv preprint arXiv:2503.06955 (2025)

    Pith/arXiv arXiv 2025