Pith. sign in

REVIEW 3 major objections 5 minor 38 references

UniMoFlow: Grounding Instruction-Driven 3D Human Motion Editing in Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper claims that grounding 3D motion editing in generation via a unified flow-matching model with full source context improves target alignment, edit effectiveness, and cycle consistency over prior editors while keeping generation…

desk verdict A serious, well-ablated unification of generation and editing for 3D human motion, with a large synthetic dataset—but the headline editing gains rest on a self-referential benchmark and need independent validation. read the letter →

arxiv 2608.09143 v1 pith:XUR3CO3S submitted 2026-08-10 cs.CV

Yilei Hua , Beibei Jing , Ce Zheng , Hanyu Zhou , Yawei Luo , Wei Yang This is my paper · ORCID
classification cs.CV
keywords 3Dhumanmotioneditingtext-to-motiongenerationinstruction-drivenflowmatchinglatenttransformersyntheticdatasetsource-anchoredsamplingevaluationmetrics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that instruction-driven editing of 3D human motion can be made a native capability of a text-to-motion generator instead of a separate trained editor. To do this it builds Omni-MoEdit, a 55,641-pair synthetic dataset in which a language model turns motion captions into edit instructions, a text-to-motion DiT synthesizes target motions through FlowEdit, and a frozen evaluator filters for target-text alignment and source preservation. Its model UniMoFlow is a latent flow-matching transformer—a generative model that learns a velocity field from noise to data—that places edit text, noisy target tokens, and the full source-motion token sequence in one self-attention context, with mode-specific masks switching between generation and editing. A source-anchored sampling mode, SAFE, edits by integrating the difference between instruction-conditioned and null-text velocity fields while starting from the source latent. If the paper is right, one compact model can generate and edit human motion, and large synthetic triplets can replace scarce manual edit annotations.

What carries the argument

The central object is a unified latent flow-matching transformer with in-context conditioning. Source and target motions are compressed by a causal VAE—a temporal autoencoder that turns every four motion frames into one continuous latent token—and a frozen T5 encoder supplies text tokens; at editing time the context sequence is noisy target, edit text, separator, and full source, with time-conditioned additive modality tags and a mode-specific global self-attention mask letting the same backbone act as generator or editor. The model is trained with rectified-flow interpolation in a length-dependent shifted time, with separate losses on edit triplets and generation samples, and text dropout that retains the source condition, which produces the null-text velocity field SAFE exploits at inference. SAFE starts from the source latent and takes steps along gated differences between instruction-conditioned and null-text velocities, with a strength parameter that trades source fidelity against edit magnitude without retraining. The accompanying Omni-MoEdit pipeline is the other load-bearing mechanism: a language model generates commands and target captions, a base DiT synthesizes targets via FlowEdit, and thresholds on Matching Score, R@1, DeltaMatch, and Structure filter the triplets.

What would settle it

Take a random sample of 500 Omni-MoEdit test triplets stratified by the five edit types and have independent human raters judge whether the synthesized target motion both realizes the edit instruction and preserves all unspecified source content; if the evaluator's acceptance thresholds (Matching Score ≥ 0.6, R@1 ≥ 0.7, DeltaMatch ≥ 0.1, Structure ≥ 0.4) pass pairs that human raters consistently reject, or if agreement differs by edit type, the dataset and all editing metrics built on that evaluator inherit the bias.

Watch

Extended reading notes

Core claim

The central claim is that grounding editing in generation works across data, architecture, and inference: a model trained jointly on text-to-motion generation and Omni-MoEdit edit triplets learns to localize and apply body-part, amplitude, temporal, action, and style changes while preserving unedited structure. UniMoFlow (SAFE) raises target-text retrieval at rank 1 from 0.4839 to 0.6347 and improves edited-vs-ground-truth retrieval from 0.5197 to 0.6762 relative to the strongest prior editor in its comparison, while cutting FID from 26.06 to 12.45; UniMoFlow also reports the best FID (15.331) and Matching Score (0.716) among continuous-space generators on the SnapMoGen test split. The authors interpret these results as showing that full token-level source access plus a shared generative prior outperforms compressed source conditioning and training-free adaptation, and that native editing and source-anchored sampling cover complementary regimes, with SAFE suited to localized refinements and the native editor more reliable for large structural reconstructions.

Load-bearing premise

The pipeline treats the frozen evaluator's similarity scores as ground truth for both 'the target motion realizes the instruction' and 'the source is preserved', so if that embedding space is biased or blind to some edit categories, the dataset and every reported editing comparison inherit the bias.

Editorial extensions

If this is right

  • Large-scale instruction-driven motion editing can be supervised without manually curated triplets: the synthesis-and-filter recipe yields 55,641 pairs covering body-part, amplitude, temporal, action, and style edits, about 3.5 times larger than prior adjustment sets.
  • Unified generation-editing training is mutually beneficial: generation batches regularize residual artifacts in synthetic edits, and edit batches strengthen language-motion grounding, so a single checkpoint serves both tasks.
  • Source-anchored inference gives a continuous edit-strength dial: adjusting the drift weight controls how far the output moves from the source toward the instruction, without retraining, which suits interactive animation workflows.
  • Semantics-aware metrics matter: reporting positive improvement ratio, region alignment, and cycle consistency alongside reconstruction metrics separates realizing the requested change from preserving the source, so under-editing no longer masquerades as success.
  • Edits that require drastic structural reconstruction remain the hard regime, with the paper stating that the native editor is more reliable there, while SAFE is preferred for fine-grained adjustments.

Reading between the lines

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

  • As an extension beyond the paper, the editing-quality ceiling is likely set by the caption diversity and the evaluator's embedding, so swapping the source corpus or the evaluator would change which edit types succeed.
  • As an extension beyond the paper, any flow-matching generator trained with source-conditioned null-text dropout could be turned into an editor with the SAFE recipe, which suggests a direct transfer test to image or audio editing.
  • As an extension beyond the paper, the reverse instructions in Omni-MoEdit could serve as a cycle-consistency training loss, not just an evaluation metric.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes UniMoFlow, a unified latent flow-matching model that handles both text-to-motion generation and instruction-driven motion editing in a single self-attention context, together with a synthetic dataset Omni-MoEdit and an inference-time refinement scheme SAFE. The dataset is built by using Qwen3-8B to generate edit instructions and target descriptions from SnapMoGen captions, synthesizing target motions with a base DiT and FlowEdit, and filtering candidates with a frozen SnapMoGen evaluator. UniMoFlow is trained jointly on generation data and Omni-MoEdit triplets, with source-motion tokens placed in the same attention context as text and noisy target tokens. SAFE initializes from the source latent and uses the difference between instruction-conditioned and null-text velocity fields to steer the edit. Experiments report substantial improvements over supervised and training-free editing baselines on target-text alignment, source preservation, edit localization, and cycle consistency, while maintaining competitive text-to-motion generation quality on the SnapMoGen test split.

Significance. If the central claim is supported, the paper makes a useful contribution: it offers a principled way to scale editing supervision from synthetic triplets, a compact architecture that unifies generation and editing, and a controllable inference-time editing mechanism. The manuscript is also strong on experimental hygiene in several respects: the baselines are reproduced from released code with a common data protocol, the ablations isolate the contributions of mixed training, latent flow matching, token-level tags, and attention masks, and the new metrics are defined explicitly in the supplementary material. The generation results in Table 4 are evaluated on an independent SnapMoGen test split and are not affected by the editing-evaluator concern. However, the main editing comparison is currently evaluated with the same frozen SnapMoGen evaluator that was used to filter the synthetic training data, and UniMoFlow is trained in the same latent space and flow-matching paradigm as the synthesis backbone. This creates a real risk that the headline gains in Tables 2 and 3 measure proximity to the synthesis pipeline rather than instruction-following.

major comments (3)
  1. [Dataset synthesis and evaluation protocol (Fig. 2, Eqs. 9–29)] The headline editing comparison is a closed loop. Candidate triplets are retained only when the frozen SnapMoGen evaluator reports Match ≥ 0.6, R@1 ≥ 0.7, Δmatch ≥ 0.1, and Struct ≥ 0.4; the same evaluator's embedding space supplies every editing metric in Tables 2 and 3 through Eqs. (9)–(29); and UniMoFlow is trained on the retained targets in the same Wan-VAE latent space and flow-matching paradigm used by the synthesis backbone. The large SAFE/Native gains over MotionLab, TMED, and OmniME may therefore reflect proximity to the SnapMoGen base-DiT target distribution rather than genuine instruction-following. The baselines are additionally disadvantaged because they operate in different motion representations (e.g., 28-joint windows for MotionReFit and raw 296-D features for MotionLab), so they cannot benefit from the same latent-space prior. The generation results in Table 4 are not affected by this concern, but the central claim that UniMoFlow 'substantially outperforms prior editors' is not yet supported. I request at least one genuinely independent validation: human preference ratings on a sampled test set, an off-the-shelf text–motion retriever not used in filtering, or a zero-shot probe on MotionFix-style human-annotated pairs. The conclusion's own concession that the system 'still depends on automatic filtering' makes this a required validation rather than an optional extra.
  2. [SAFE drift strength (Table 7) and generation CFG (Fig. 7)] The main SAFE results use w = 1.0, which is selected from the sweep in Table 7 on Omni-MoEdit with no named validation split. If this sweep was performed on the test split that is also used for Tables 2 and 3, the reported SAFE numbers are the best of five hyperparameter settings and the comparison is optimistic. Please either perform the w sweep on the held-out validation split and freeze w before test evaluation, or report Tables 2 and 3 for a range of w values (e.g., 0.5, 1.0, 2.0) so the reader can assess the sensitivity of the main claims. A similar issue applies to the generation results: the CFG scale 4.5 in Table 4 is selected on the SnapMoGen test split according to Figure 7. The generation comparison should use a validation-based selection, or the authors should justify why test-set selection does not bias the comparison against the baselines.
  3. [SAFE mechanism (Eqs. 7–8)] The paper states that the residual v_c − v_u 'distills the velocity component attributed to the edit instruction while remaining anchored to the source motion.' This is an important claim for SAFE, but it is not formally established or empirically isolated. Table 7 shows that larger w degrades FID and retrieval, which is consistent with drift away from the source manifold, but there is no analysis of how the residual behaves across edit types, motion lengths, or source-motion complexity. I would like to see a quantitative decomposition, for example measuring source fidelity and edit success separately for localized edits versus structural edits, to support the claim that SAFE is complementary to native editing rather than simply a weaker version of it.
minor comments (5)
  1. [Introduction and Related Work] The sentence 'For zero-shot or training-free editing, we mainly use such samplers to synthesize candidate pairs with Qwen3-8B' is confusing about whether training-free editors are used only for data synthesis or also as evaluation baselines; please clarify.
  2. [Equation (5)] Equation (5) has unbalanced delimiters in the expectation terms; please fix the typesetting.
  3. [SAFE section] The paragraph on source-anchored null-text fields ends with 'Section .' and a blank cross-reference; please fill in the correct section number.
  4. [Tables 2 and 3] The abbreviations 'UniMoFlow (S)' and 'UniMoFlow (SAFE)' are used inconsistently across Tables 2 and 3; please unify them with the main text.
  5. [Evaluation metric definitions (Eq. 16)] The statement that 'only the first 148 motion channels are used' for the Struct metric requires a citation to the evaluator implementation or a specification of which channels are included; otherwise the metric cannot be reproduced independently.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the editing benchmark is self-referential in construction, but no predicted result reduces to a fitted input by the paper's own equations.

full rationale

The paper contains no circular derivation in the formal sense used by this review. The central claims are empirical: UniMoFlow is trained with the flow-matching objective in Eq. 5 on Omni-MoEdit triplets, and its editing performance is compared against re-trained baselines on the same train/validation/test partition. No equation in the paper defines a predicted quantity in terms of the fitted parameters or of the benchmark's acceptance thresholds. The filtering thresholds (Match>=0.6, R@1>=0.7, DeltaMatch>=0.1, Struct>=0.4) and the evaluation metrics (Eqs. 9-30) both use the frozen SnapMoGen evaluator, which makes the editing benchmark self-referential as a measure of real-world instruction following and may bias scores toward models that reproduce the SnapMoGen target distribution; UniMoFlow shares the Wan-VAE latent space and flow-matching paradigm with the synthesis backbone. However, this is a dataset-construction and external-validity limitation, not a circular step: the compared methods are all trained and scored under the same protocol, the model is not trained to optimize the evaluator directly, and the text-to-motion generation results in Table 4 are measured on the independent SnapMoGen test split against external baselines such as T2M-GPT, MoMask, and MDM. All external dependencies (SnapMoGen, FlowEdit, Qwen3-8B, Wan-VAE, UMT5) are prior work by other groups; there is no load-bearing self-citation or imported uniqueness theorem. The conclusion explicitly concedes dependence on automatic filtering and defers human preference signals to future work. Accordingly, no specific circular step can be quoted, and the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on a chain of modeling and evaluation assumptions rather than on derived mathematics. The main load-bearing elements are: (1) the SnapMoGen evaluator is a valid proxy for edit correctness and is used both to filter the dataset and to score all editing metrics; (2) FlowEdit plus a SnapMoGen-trained DiT produces valid target motions; (3) Qwen3-8B text generation is reliable; (4) the causal VAE preserves sufficient detail; and (5) SAFE's residual velocities behave as an edit-direction field. No new physical entities are introduced. Several hand-set hyperparameters (filtering thresholds, SAFE drift w, CFG scale 4.5, loss balance) directly affect the reported numbers.

free parameters (5)
  • Omni-MoEdit filtering thresholds = Match >= 0.6, R@1 >= 0.7, DeltaMatch >= 0.1, Structure >= 0.4
    Hand-set thresholds in data synthesis (Section 'The Omni-MoEdit Dataset', Figure 2) determine which triplets are retained; no sensitivity analysis or validation-based selection is reported.
  • SAFE drift strength w = 1.0
    Selected from Table 7 sweep on Omni-MoEdit without stating validation split; all main SAFE editing results use this value.
  • Classifier-free guidance scale for generation = 4.5
    Chosen from Figure 7 FID sweep on SnapMoGen test split; main generation results use this value.
  • Loss weight between generation and editing = lambda_g = lambda_e (equal weights)
    Training configuration states equal weights without an ablation of the mixing ratio, though mixed training itself is ablated.
  • Length-dependent time shift s(t;l)
    Equation 4 introduces a monotone reparameterization depending on sequence length, but the concrete functional form is not specified in the main text; the model depends on it.
assumptions (6)
  • standard math Rectified flow interpolation z_t = (1-t)e + t z and the chain-rule velocity v* = (z - e)/J(t;l) are valid generative training targets.
    Equations 3 and 4 define the flow-matching objective; this is standard math for rectified flows, not a domain-specific assumption.
  • domain assumption The frozen SnapMoGen evaluator's embedding similarities are a valid proxy for edit fidelity, target-text alignment, and source preservation.
    Used to filter Omni-MoEdit pairs (thresholds 0.6, 0.7, 0.1, 0.4) and to compute every editing metric in Tables 2 and 3; if this proxy is biased, the benchmark is biased.
  • domain assumption FlowEdit on a SnapMoGen-trained base DiT can synthesize target motions that realize action, timing, amplitude, and style edits with acceptable quality.
    The whole Omni-MoEdit dataset depends on using FlowEdit as target synthesizer; no independent human validation of synthesized targets is reported.
  • domain assumption Qwen3-8B-generated edit instructions, target captions, and reverse instructions are semantically correct and paired consistently with the source caption.
    Data pipeline (Figure 2) uses Qwen3-8B for all triplet text; errors here propagate into training and cycle-consistency evaluation.
  • ad hoc to paper SAFE's source-anchored velocity residual v_c - v_u isolates the instruction-attributed edit while staying on the source motion manifold.
    Equation 8 defines SAFE; the paper assumes this residual is the meaningful edit direction and that gated drift controls the edit-preservation tradeoff. This is the paper's own modeling hypothesis.
  • domain assumption A 32-D, 4x temporally compressed causal VAE latent preserves the kinematic detail needed for both generation and editing.
    All motions are encoded to this latent; Table 6 ablates latent dimension but does not establish that 32-D is sufficient for fine-grained temporal edits.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniMoFlow: Grounding Instruction-Driven 3D Human Motion Editing in Generation." pith.science (2026). https://pith.science/paper/XUR3CO3S

@misc{pith2026260809143,
  author       = {Pith},
  title        = {Pith review of: UniMoFlow: Grounding Instruction-Driven 3D Human Motion Editing in Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XUR3CO3S}},
  note         = {Machine review of arXiv:2608.09143}
}
read the original abstract

Instruction-driven editing of 3D human motion requires precise spatiotemporal localization, rich semantic grounding, and strict preservation of unmodified content. Existing methods either resort to training-free adaptation of generative models or rely solely on triplet supervision; however, adaptation often yields suboptimal control, and manually curated triplet datasets remain severely limited in scale and semantic diversity. To overcome this bottleneck, we ground motion editing directly within text-to-motion generation across data, architecture, and inference. At the data level, we develop a closed-loop synthesis-and-verification pipeline that produces Omni-MoEdit, a large-scale dataset spanning body-part, amplitude, temporal, action, and style edits. At the architectural level, we introduce UniMoFlow, a unified latent flow-matching model that shares broad semantic and kinematic knowledge between generation and editing. At the inference level, SAFE (Source-Anchored Flow Editing) complements UniMoFlow with controllable, source-anchored refinement. Furthermore, we augment standard evaluations with semantics-aware metrics to account for valid edits that inherently deviate from a single ground-truth reference. Extensive experiments demonstrate improved target-text alignment, edit effectiveness, and cycle consistency, while maintaining competitive source fidelity and text-to-motion generation quality.

Figures

Figures reproduced from arXiv: 2608.09143 by the authors.

Figure 1
Figure 1. We propose UniMoFlow, a unified method for high-quality motion generation and editing. Given a generated or user-provided source motion sequence, UniMoFlow adjusts body-part movements, action types, timing, amplitude, style, and compound edits while faithfully following textual instructions. Abstract Instruction-driven editing of 3D human motion requires pre￾cise spatiotemporal localization, rich semantic grounding,… view at source ↗
Figure 2
Figure 2. The Omni-MoEdit data synthesis pipeline. The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Architecture of UniMoFlow. Text, noisy target-motion, and optional full source-motion tokens form a unified context [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: SAFE inference. Starting from the source latent, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison on Omni-MoEdit. Red and orange boxes highlight motion regions that satisfy the edit-command semantics. edit. For text-to-motion generation, R-Precision Top-3 and Matching Score measure text-motion alignment, FID mea￾sures distribution quality, an…
Figure 6
Figure 6. Figure 6: Composition of Omni-MoEdit. (a) Overlapping frequencies of the five basic edit types. (b) Number of edit types [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Generation FID under different classifier-free guid [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Additional qualitative comparisons on the Omni-MoEdit test split. Each row applies the same source motion and edit [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Five isolated UniMoFlow edits from one forward-walking source, covering action type, body part, amplitude, timing, [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Representative failures: an abstract continuation after dense soccer motion (row 1), a right-hand edit conflicting with [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 33 canonical work pages

  1. [1]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Generating Diverse and Natural 3D Human Motions from Text , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    work page

  2. [2]

    and Varol, G

    Petrovich, Mathis and Black, Michael J. and Varol, G. Proceedings of the European Conference on Computer Vision , pages =

    work page

  3. [3]

    2024 , doi =

    Zhang, Mingyuan and Cai, Zhongang and Pan, Liang and Hong, Fangzhou and Guo, Xinying and Yang, Lei and Liu, Ziwei , journal =. 2024 , doi =

    work page 2024

  4. [4]

    International Conference on Learning Representations , year =

    Human Motion Diffusion Model , author =. International Conference on Learning Representations , year =

    work page

  5. [5]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Executing Your Commands via Motion Diffusion in Latent Space , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    work page

  6. [6]

    Guo, Chuan and Zuo, Xinxin and Wang, Sen and Cheng, Li , booktitle =

    work page

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Generating Human Motion from Textual Descriptions with Discrete Representations , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    work page

  8. [8]

    Jiang, Biao and Chen, Xin and Liu, Wen and Yu, Jingyi and Yu, Gang and Chen, Tao , booktitle =

    work page

Show all 38 references
  1. [9]

    Guo, Chuan and Mu, Yuxuan and Javed, Muhammad Gohar and Wang, Sen and Cheng, Li , booktitle =

  2. [10]

    Pinyoanuntapong, Ekkasit and Wang, Pu and Lee, Minwoo and Chen, Chen , booktitle =

  3. [11]

    and Varol, G

    Athanasiou, Nikos and Ceske, Alpar and Diomataris, Markos and Black, Michael J. and Varol, G. SIGGRAPH Asia 2024 Conference Papers , year =

    2024

  4. [12]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Dynamic Motion Blending for Versatile Motion Editing , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  5. [13]

    Guo, Ziyan and Hu, Zeyu and Soh, De Wen and Zhao, Na , booktitle =

  6. [14]

    Li, Zhengyuan and Cheng, Kai and Ghosh, Anindita and Bhattacharya, Uttaran and Gui, Liangyan and Bera, Aniket , booktitle =

  7. [15]

    Yang, Yujie and Zhang, Zhichao and Chen, Jiazhou and Wu, Zichao , journal =

  8. [16]

    Bu, Wendong and Pan, Kaihang and Lin, Yuze and Li, Jiacheng and Shen, Kai and Zhang, Wenqiao and Li, Juncheng and Xiao, Jun and Tang, Siliang , journal =

  9. [17]

    and Sridhar, Srinath , journal =

    Cong, Xiaoyan and Li, Zekun and Dou, Zhiyang and Li, Hongyu and Taheri, Omid and Guo, Chuan and Mittal, Abhay and An, Sizhe and Komura, Taku and Matusik, Wojciech and Black, Michael J. and Sridhar, Srinath , journal =

  10. [18]

    Kulikov, Vladimir and Kleiner, Matan and Huberman-Spiegelglas, Inbar and Michaeli, Tomer , booktitle =

  11. [19]

    Hong, Seokhyeon and Lee, Chaelin and Han, Mingi and Jang, Jaehyung and Lee, Junhyeong and Kim, Jinwoo and Oh, Tae-Hyun , booktitle =

  12. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Omni-Supervised Motion Editing: Balancing Change and Invariance through Positive-Negative Learning , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  13. [21]

    Jiang, Nan and Li, Yunhao and Pang, Lexi and He, Zimo and Huang, Siyuan and Zhu, Yixin , booktitle =

  14. [22]

    arXiv preprint arXiv:2410.18977 , year =

    Pay Attention and Move Better: Harnessing Attention for Interactive Motion Generation and Training-free Editing , author =. arXiv preprint arXiv:2410.18977 , year =

    arXiv

  15. [23]

    Guo, Chuan and Hwang, Inwoo and Wang, Jian and Zhou, Bing , booktitle =

  16. [24]

    Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Gao, Chang and Huang, Chengen and Lv, Chenxu and Zheng, Chujie and Liu, Dayiheng and Zhou, Fan and Huang, Fei and Ge, Hao and Wei, Haoran and Lin, Huan and Tang, Jialo...

  17. [25]

    , booktitle =

    Brooks, Tim and Holynski, Aleksander and Efros, Alexei A. , booktitle =

  18. [26]

    arXiv preprint arXiv:2506.15742 , year =

    arXiv

  19. [27]

    Xia, Bin and Zhang, Yuechen and Li, Jingyao and Wang, Chengyao and Wang, Yitong and Wu, Xinglong and Yu, Bei and Jia, Jiaya , journal =

  20. [28]

    Alekseenko, Grigorii and Gordeev, Aleksandr and Tolstykh, Irina and Suleimanov, Bulat and Dokholyan, Vladimir and Fedorov, Georgii and Yakubson, Sergey and Tsybina, Aleksandra and Chernyshov, Mikhail and Kuprashevich, Maksim , journal =

  21. [29]

    Han, Zhen and Jiang, Zeyinzi and Pan, Yulin and Zhang, Jingfeng and Mao, Chaojie and Xie, Chenwei and Liu, Yu and Zhou, Jingren , journal =

  22. [30]

    Wu, Chenyuan and Zheng, Pengfei and Yan, Ruiran and Xiao, Shitao and Luo, Xin and Wang, Yueze and Li, Wanli and Jiang, Xiyan and Liu, Yexin and Zhou, Junjie and Liu, Ze and Xia, Ziyi and Li, Chaofan and Deng, Haoge and Wang, Jiahao and Luo, Kun and Zhang, Bo and Lian, Defu and...

  23. [31]

    Liu, Shiyu and Han, Yucheng and Xing, Peng and Yin, Fukun and Wang, Rui and Cheng, Wei and Liao, Jiaqi and Wang, Yingming and Fu, Honghao and Han, Chunrui and Li, Guopeng and Peng, Yuang and Sun, Quan and Wu, Jingwei and Cai, Yan and Ge, Zheng and Ming, Ranchen and Xia, Lei an...

  24. [32]

    Cai, Qi and Chen, Jingwen and Chen, Yang and Li, Yehao and Long, Fuchen and Pan, Yingwei and Qiu, Zhaofan and Zhang, Yiheng and Gao, Fengbin and Xu, Peihan and Wang, Yimeng and Yu, Kai and Chen, Wenxuan and Feng, Ziwei and Gong, Zijian and Pan, Jianzhuang and Peng, Yi and Tian...

  25. [33]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Scalable Diffusion Models with Transformers , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  26. [34]

    arXiv preprint arXiv:2503.20314 , year =

    arXiv

  27. [35]

    2026 , eprint =

    A Unified Conditional Flow for Motion Generation, Editing, and Intra-Structural Retargeting , author =. 2026 , eprint =

    2026

  28. [36]

    Proceedings of the 32nd ACM International Conference on Multimedia , year =

    StableMoFusion: Towards Robust and Efficient Diffusion-based Motion Generation Framework , author =. Proceedings of the 32nd ACM International Conference on Multimedia , year =

  29. [37]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Rethinking Diffusion for Text-Driven Human Motion Generation: Redundant Representations, Evaluation, and Masked Autoregression , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  30. [38]

    arXiv preprint arXiv:2403.03206 , year =

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis , author =. arXiv preprint arXiv:2403.03206 , year =

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.