acceptodds
Under review as a conference paper at ICLR 2027

Learning Style, Reusing Motion: Prior-Guided Unpaired Motion Stylization

Abstract

Motion stylization controls how a text-specified motion is performed while preserving its content. Paired performances of the same content in different styles are scarce, limiting direct supervision. Pretrained text-to-motion generators offer broad content capabilities and motion knowledge, but style edits may conflict with both. Our framework, TriPrior, coordinates two data priors from unpaired motions with the frozen generator's own motion prior. These priors act at inference time by correcting flow-model velocity predictions or, equivalently, diffusion-model score predictions. The dataset-level prior supplies a stable style direction, the instance-level prior adapts it to the generated motion, and the generative prior provides content-conditioned correction through the generator's response to the edited sampling state. Keeping the data priors outside the generator enables their reuse on HY-Motion, HY-Lite, and Kimodo without backbone retraining. On HumanML3D with style-motion input, TriPrior on HY-Motion achieves 29.40% average accuracy across four independent style classifiers (SRA-4) and 38.70% content-retrieval R@1, exceeding MoMo and MLD+Aberman by 3.06 and 11.32 percentage points, respectively. Controlled ablations show that the generative prior improves content retrieval and motion quality while largely preserving style recognition across the evaluated styles and content motions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
motion generation, motion stylization, unpaired learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_a8SBOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addressed the problem of style editing. Previously, text-to-motion generators usually suffer from limited ability on style editing. In this paper, a new framework called, TriPrior, is presented, which coordinates two data priors from unpaired motions with the frozen generator's own motion prior. Experiments on HumanML3D and HY-motion show reasonable performance gain over the baselines. Also, a set of ablations are reported to validate the setup of experiments.

Critical strengths
  • The problem of motion stylization is interesting and useful in industry applications. Thus, the contribution of this topic is meaningful to the community.

  • The paper presents an extensive ablations to show the performance on different setups.

  • The proposed idea is easy to implement reproduce.

Critical weaknesses
  • The generalization ability of the proposed algorithm should be validated. Currently, the experiment is reported on the HumanML3D dataset with 100 style references. How about the generalizaiton ability to unseen-style or multi-style composition?

  • According to the results shown in Table 1, the performance of the proposed algorithm on HumanML3D does not have obvious performance gain and the absolute value of the result like R@1 may not meet the requirement of industry applications.

  • How about the performance on more benchmarks like AMASS or AIST++?

  • The main table claims three seeds but reports means only, and ablations use a single seed. It would be better to run multiple times and the variance is reported in the paper as the differences such as 29.18 vs 29.22 are uninterpretable.

Questions for authors

Please mainly address the questions in the weakness section. Besides, there are some minor questions.

  • How are T_init / T_refine boundaries and ratios chosen, and how sensitive are results to them?

  • How about the analysis of failure cases.

1 comment

Could you please take a look and let us know if you have any further concerns?

Reviewer_UXDoOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

TriPrior stylizes frozen text-to-motion generators using three inference-time priors: a content-balanced target-minus-Neutral dataset direction, motion-specific classifier guidance, and the frozen generator’s response to a proposed edit. Clean-motion edits are converted to flow velocities or diffusion scores; shared learned modules are reused on HY-Motion, HY-Lite, and Kimodo with representation and strength adaptations. A 4,196-task HumanML3D/100STYLE protocol over three generation seeds reports independent four-classifier style accuracy, content retrieval, distribution and contact diagnostics. On HY-Motion with style-motion input, SRA-4 is 29.40% and R@1 is 38.70%. Stage and component ablations, strength sweeps, denoising recovery tests, and a small blinded user study support a style–content–quality trade-off.

Critical strengths

S1:External style control is reused across substantially different frozen generators, and evaluation separates guidance from scoring. The same auxiliary modules transfer to HY-Lite and Kimodo; four heterogeneous frozen evaluators, common task manifests, and three complete generation cohorts provide useful evidence of reusable style knowledge rather than merely backbone-specific classifier optimization. S2:The current experiments meaningfully separate generator response from later repair. Table 3 shows sampler-stage FID 2.342→2.089 and Skate 10.82→9.41 before post-processing, while Table 4 and Figure 5 expose the style–quality cost. Removing post-processing nearly doubles Skate (4.56→8.73) without materially changing style. These distinctions make the results more informative and avoid attributing all physical gains to a single component.

Critical weaknesses

W1:The novel value of the generator response is not fully distinguished from shrinking the proposed style edit. Supplement Eq. 8 contains an explicit [1-β(1-t)]δ shrinkage term even without a network response. The full method loses 3.02 SRA-4 points versus no generative prior while improving other metrics; Figure 5 is useful, but is not a direct norm- or style-matched shrinkage-only control. The main-paper average content benefit also appears after calibration rather than at the raw sampler (Table 3 R@1 37.70 off versus 37.49 on). This narrows which improvements can be attributed specifically to learned generative dynamics. W2:The reusable-controller claim is supported, but its practical scope is narrower and more costly than the headline may suggest. References are mapped to one of 100 classes rather than used as continuous instance-specific style codes; Kimodo requires coordinate conversion and tailored strengths. Supplement Table 11 reports 14.85 versus 1.24 s/motion for HY and 13.00 versus 0.28 for Kimodo, approximately 12× and 46× plain generation latency. The eight-item user study is too small to establish perceptual generality across 100 styles. These limitations matter for impact and should be stated prominently, though they do not erase the demonstrated transfer.

Questions for authors

Q1:On the existing HY tasks, can you compare generator response with shrinkage-only updates at comparable correction magnitude, keeping acceptance checks and repair fixed and reporting sampler/final outputs? Please explain how this mechanism adds to clean-prediction guidance in DPS and Universal Guidance. Q2:What are the exact encoders/checkpoints used for acceptance checks and R@1/MM/FID evaluation? If they share an encoder, does the content advantage persist under an independent evaluator or a limited disabled-ranking control? This asks about possible coupling, not established leakage. Q3:Please report held-out real-motion accuracy for the four SRA-4 classifiers and whether evaluation references were seen during recognizer/guidance/bank training. On existing outputs, how do results differ for correctly and incorrectly classified references, and is reference information beyond the closed-set style label used? Q4:How are mirror pairs in the 4,196 tasks handled in uncertainty estimates? Please give the user-study participant count and item-level preferences, and reconcile the content-improvement claim with Table 3’s slight sampler R@1 decrease when response is enabled. Q5:Beyond the existing response/post-processing comparison and denoising-recovery test, can existing outputs distinguish the final gains from real-motion calibration, denoising and contact repair individually? Please keep other settings fixed or explain inseparable stage dependencies. Q6:For a new backbone, which modules/statistics transfer unchanged and which conversions, strengths or acceptance rules require adaptation? How were settings selected without test-set tuning? Given Table 11’s reported latency/memory, please explain the full-pipeline overhead and any available stage-wise breakdown.

1 comment

Could you please let us know if our responses have addressed your concerns?

Reviewer_fJCROpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TriPrior, which adds style control to existing text-to-motion models without retraining the generators. It starts with a style direction estimated from real motions, then refines the result using a style classifier and feedback from the generator. The authors reuse the same learned style modules across HY-Motion, HY-Lite, and Kimodo, with backbone-specific representation conversions and sampling settings. Experiments on HumanML3D and 100STYLE show promising results in expressing the target style while preserving the requested motion content.

Critical strengths
  1. Reuse across generators is practically useful. The same learned style components work with HY-Motion, HY-Lite, and Kimodo without training a new style branch for each backbone. This is a useful advantage as stronger motion generators become available.
  2. The experiments support the system’s practical value. The shared evaluation protocol, three generation seeds, and separate style evaluators provide useful support for the reported results. The ablations also help explain the trade-off between style expression and content preservation.
Critical weaknesses
  1. The contribution of the generator response is not fully isolated. Appendix Eq. (8) combines edit attenuation with a velocity-field response. Figure 5 provides useful guidance-strength comparisons, but an attenuation-only control is still missing. This leaves some uncertainty about how much the additional generator query contributes.
  2. The evaluation scope remains narrow relative to the motivation for backbone reuse. Experiments use HumanML3D descriptions and 100STYLE references. Testing broader prompts from HY-Motion and Kimodo, along with compatible Kimodo stylized locomotion references, would better establish whether the controller preserves the broader capabilities of these generators.
Questions for authors
  1. Contribution of the generator response. Figure 5 provides useful comparisons across guidance strengths. Could you also compare the full HY-Motion update with an attenuation-only control that retains the shrinkage term in Appendix Eq. (8) but removes the velocity-field response? Keeping the correction caps, acceptance checks, and post-processing fixed, and reporting results before and after post-processing, would help isolate the benefit of the generator response.
  2. Content-ranking acceptance check. Which text–motion encoder and checkpoint are used for the acceptance check in Appendix B.2? Is this the same encoder used to compute R-precision and MM? An ablation of this check, with the other components fixed, would help clarify its contribution to content preservation.
  3. Evaluation beyond HumanML3D and 100STYLE. Could you provide a small evaluation using prompts from HY-Motion and Kimodo, with the controller and existing settings for each backbone kept fixed? I would also like to see some Kimodo stylized locomotion motions used as references, choosing styles covered by 100STYLE and reporting their predicted classes. These tests would strengthen the evidence that the controller handles broader motion content and references from another source.

Comments on the paper

1 comment
Area_Chair_NNMpOpenReview ↗

Read author feedback and discuss

Dear Reviewers,

Thanks for reviewing this paper. Please take a closer look at the feedback from the authors and see if it has addressed your concerns and if you have follow-up questions for the authors. The discussion session will end soon!

Best, AC for paper 5348

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.