acceptodds
Under review as a conference paper at ICLR 2027

Controllable Exaggeration for Generative Motion Models via Training-Time Adaptation and Inference-Time Guidance

Abstract

Recent motion generative models have demonstrated strong capabilities in synthesizing physically plausible character motion, but often overlook established animation principles used by professional animators to ground and design their animation work. Understanding and incorporating these principles into motion generative pipelines is essential for producing motions that serve not only physically grounded applications but also the needs of the character animation community. This enables the creation of characters that not only move in physically plausible ways but also feel alive, expressive, and engaging. To close this gap, we focus on the Exaggeration principle of animation and investigate how it can be incorporated into modern motion generative pipelines to produce more expressive character motions. To this end, we introduce a framework that operates at two stages of existing motion generative pipelines. The first stage introduces exaggeration during training, where we perform supervised fine-tuning of pre-trained text-to-motion models on our curated exaggeration dataset. The second stage operates at inference time, where we: (i) introduce a mathematical formulation of exaggeration based on dynamic movement primitives (DMPs); and (ii) leverage this formulation as an exaggeration guidance signal to guide existing diffusion and flow-matching text-to-motion generation models toward exaggerated motion without additional training. Through qualitative and quantitative evaluations against three strong motion generation models, we show that our methods generate more exaggerated and expressive motions while preserving neutral reference motion intent and physical plausibility.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Motion Generative Models, Training-Free Guidance, Motion Exaggeration, Supervised Fine-Tuning, Animation Principles, Character Animation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_5bv3OpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies controllable motion exaggeration in pretrained text-to-motion models, aiming to increase expressiveness while preserving action intent and physical plausibility. The authors collect ExMoCap, a dataset of 26 actions at five exaggeration levels, and investigate two complementary approaches: LoRA fine-tuning with prior preservation, and inference-time guidance toward motion targets produced by scaling DMP forcing terms. The approaches are evaluated individually and in combination on MDM / HY-Motion / KIMODO, using quantitative metrics, qualitative examples and a user study.

Critical strengths
  1. Graded motion-capture data for studying controllable exaggeration. ExMoCap covers 26 actions performed at five exaggeration levels, providing direct supervision for learning intensity control while keeping the action category fixed. This structured collection is a useful empirical contribution to the study of expressive character animation.

  2. Promising results for training-time exaggeration control. The evaluation covers both diffusion and flow-matching backbones. In particular, SFT on HY-Motion-1B improves the reported ELRA@1 from 0.354 to 0.938, while R-precision@3 changes only from 0.7422 to 0.7383 (Table 1). These results suggest that pretrained motion generators can learn graded exaggeration with limited loss of text–motion alignment.

Critical weaknesses
  1. Limited methodological advance over closely related work. The combination of LoRA adaptation and prior preservation for motion generation is already central to Dance Like a Chicken. Scaling DMP forcing terms to produce exaggeration was explored by Hielscher et al., as explicitly acknowledged in Sec. 3.3.1. The broader strategy of modifying clean-motion predictions during sampling also has precedents in GMD. At the task level, Deep Motion Warping via Phase-Conditioned Diffusion Autoencoder published at SIGGRAPH 2026, already includes motion exaggeration among its editing capabilities and warrants a direct discussion. The multi-keyframe formulation and its integration with pretrained motion generators may offer useful extensions, but the current comparisons do not establish their specific advantages over existing adaptation and editing approaches.

  2. The benefit of inference-time guidance is not established at matched exaggeration levels. Table 1 shows improvements in motion-quality metrics relative to direct DMP editing alongside reductions in exaggeration metrics across all three backbones. For example, on MDM, foot-skate decreases from 0.134 to 0.091, while δphys\delta_{\mathrm{phys}} decreases from 0.953 to 0.552. These results leave open whether comparable quality could be obtained simply by reducing the strength of direct DMP editing. Comparisons at matched achieved exaggeration levels, including direct DMP editing with adjusted strength and DMP followed by standard contact or kinematic correction, are needed to establish the specific benefit of iterative guidance. The user study does not resolve this comparison, and its action-preservation evaluation covers only six motion examples, five of which use KIMODO (Table 5), limiting support for the broader preservation claims.

Questions for authors
  1. Contribution relative to prior work. What is the substantive methodological advance over LoRA-MDM, the DMP-based exaggeration of Hielscher et al., and guidance approaches such as GMD? Please also clarify how the scope and capabilities differ from Deep Motion Warping (SIGGRAPH 2026). Which comparisons or targeted ablations establish the specific benefit of the multi-keyframe DMP formulation and its use as a guidance target?

  2. Comparison at matched exaggeration levels. Does TF still improve motion quality and preserve action intent when compared at the same achieved exaggeration level with direct DMP editing and DMP followed by contact or kinematic correction? Please provide comparisons across several intensities using the same neutral motions and seeds, including blinded judgments of naturalness and action preservation alongside physical-artifact measures. Since Δphys\Delta_{\mathrm{phys}} is standardized within each action, please clarify whether its normalization is shared across methods and use a common intensity scale, such as raw ROM, for matching.

  3. Evaluation independence and sample selection. Please clarify the data partitions used for SFT and for training and validating the exaggeration classifiers, including the number of original recordings in each partition. Are all windows and augmented variants from the same recording kept within one partition? What accuracy do the classifiers achieve on held-out original recordings, and which evaluated actions, if any, were excluded from SFT? Please also describe how the motions used in the quantitative evaluation and user study were selected, including the sampling procedure and any exclusions.

If these comparisons establish a consistent advantage at matched exaggeration levels and the evaluation concerns are addressed, I would update my assessment accordingly.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 5bv3 - Weakness 1 and Question 1


1- The answer to the first part - Methodological novelty over prior work

We thank the reviewer for raising this point and for giving us the opportunity to clarify how our work builds on prior methods and what it contributes. We agree that our work builds on established components, and we do not claim LoRA, prior preservation, DMPs, or clean-motion guidance individually as novel. The substantive contribution is their formulation and study for a previously underexplored problem in generative human motion: generating the same action with controllable levels of an animation principle while preserving its identity and physical plausibility.

LoRA-MDM studies few-shot transfer of a reference style to MDM. We adopt its LoRA and prior-preservation strategy, but address a different learning problem: five controlled exaggeration levels captured across 26 actions. We further study whether this control can be learned by three different motion-generation architectures. The contribution of this branch is therefore the controlled formulation, dataset, and cross-model empirical study, rather than a new LoRA algorithm.

Hielscher et al. formulate animation principles using DMPs for expressive robot motion. We extend this idea to multi-keyframe, full-body character motion and use the resulting exaggeration operator to construct targets for pretrained motion generators. The per-interval formulation and its integration with diffusion and flow-matching priors are specific to our setting.

GMD establishes the broader idea of modifying clean-motion predictions during diffusion sampling. However, it addresses externally provided spatial constraints such as root trajectories, obstacles, and sparse keyframes, using feature projection, imputation, and dense spatial guidance. Our method instead automatically derives a full-body, action-dependent exaggeration target from the model’s own neutral motion. Our claim is therefore not the generic act of modifying a clean prediction, but the DMP-based target construction and its use for controllable exaggeration within frozen motion generators.

Deep Motion Warping is a learned editing framework that requires an existing motion and a separately trained phase-conditioned diffusion autoencoder. It supports exaggeration as one of several edits by manipulating disentangled motion components. Our work instead introduces controlled exaggeration directly into pretrained text-to-motion generation, through either supervised adaptation or training-free guidance. The methods therefore differ in input, model requirements, and objective. We will add a direct discussion of this concurrent work and clarify these distinctions.

Overall, our novelty lies in the complete problem formulation: controlled exaggeration for generative motion, its training-time and training-free realization across multiple foundation models, and an evaluation suite designed to measure controllability, preservation, and quality. We have now revised the manuscript to state this contribution more precisely.


2- The answer to the second part - Clarification on multi-keyframe DMP and guidance

We thank the reviewer for asking us to clarify what each comparison establishes. The multi-keyframe formulation and the use of DMP motion as a guidance target address two distinct issues.

First, fitting a single DMP to a complete character motion sequence assumes a single start-to-goal trajectory, which cannot accurately represent actions containing multiple key poses. Specifically, our multi-keyframe formulation, as described in Sec. 3.3.1, fits a separate DMP between consecutive key poses (frames), so that each interval has its own start pose, goal pose, and fitted forcing term. This makes the intermediate poses explicit anchors for the exaggeration target. To promote smooth transitions, we initialize each successive DMP rollout using the previous interval's terminal velocity, adjusted according to the relative interval lengths. As a result, without such a multi-keyframe formulation, estimating and interpolating the motion between key poses (frames) almost completely fails across all motion examples, as demonstrated in our experiments during the study. Therefore, the multi-keyframe formulation is a necessity rather than an optional choice for properly exaggerating the entire character motion sequence in the problem we are addressing.

Second, the benefit of integrating this DMP target with a pretrained generator is established by the direct-DMP versus TF comparison in Table 1 in the manuscript. This comparison isolates the effect of adding generative guidance to the DMP target. Across all three backbones, TF substantially improves text-motion alignment, FID, and foot-skate over direct DMP exaggeration, showing that the pretrained motion prior regularizes the exaggerated target.

AuthorsOpenReview ↗

Response to Reviewer 5bv3 - Question 3

We thank the reviewer for requesting these details. We agree that the relationship between the SFT data, classifier partitions, and evaluation examples should be clearer. We provide these details and address the reviewer's questions point by point below.

SFT data preparation and partitions. Our raw dataset contains 132 distinct recordings across 26 actions. Our data pipeline constructs approximately 16 windows per recording, with fewer windows for short recordings, and creates one mirrored and one jittered copy of each window. Partitioning is performed by source window. This means each window and its augmented variants remain together in training or validation. However, different windows from the same recording can appear in both partitions. Therefore, the protocol does not enforce recording-, action-instance-, or actor-disjoint splits. This data augmentation/partitioning gives the following counts in Table 1.

Table 1. ExMoCap data partitioning and augmentation

PartitionOriginal windowsMirrored copiesJittered copiesTotal clipsOriginal recordings
Training1874187418745622132
Validation208208208624109

Of these 132 original recordings, 109 contribute different windows to both training and validation, while 23 contribute windows only to training. No recording contributes exclusively to validation. For example, a recording producing 16 windows could contribute 14 to training and two to validation; it would count as one shared recording. Each window’s mirrored and jittered variants remain in the same partition as that window. This is the answer to the reviewer’s question, “Are all windows and augmented variants from the same recording kept within one partition?”. Also, please note that the grouping preserves each window and its augmented variants together, but does not keep all windows from an original recording together.


Classifier training, validation, and performance. The classifiers use the same training partition as SFT. They are trained separately from the motion generators, remain frozen during evaluation, and are not used in the SFT objective or inference-time guidance. As for the classifier evaluation, the 624-clip validation pool, obtained in the SFT data partitioning stage above, is divided equally by exaggeration level, into an early-stopping subset and a test subset. The early-stopping subset is used to select the best classifier checkpoint and the test subset is used to measure the checkpoint final accuracy. This secondary division, which is specific only to the classifiers, operates at the individual-clip level, so augmented variants of a validation window can appear in both classifier subsets. Table 2 presents the detailed numbers of motion clips used in the validation and test subsets of the classifier.

Table 2. Classifier validation and test partitions for KIMODO.

Partition detailCount
Early-stopping (validation) clips310
Unique recordings in validation set101
Test clips314
Unique recordings in test set101

Also, the performance of the classifiers, in terms of test accuracy, is presented in Table 3.

Table 3. Performance of the classifiers trained separately for each motion-generation backbone.

ClassifierTest accuracy
MDM97.44%
HY-Motion98.73%
KIMODO98.73%

Actions excluded from SFT. No action category was deliberately withheld from SFT for the original 26-action exaggeration evaluation.


Quantitative evaluation sampling. The quantitative evaluation in Table 1 of the paper, is based on one motion for each of 26 actions and five ordered control settings, giving 130 motions per method and backbone for the exaggeration-control and foot-skate metrics. Reported Novelty uses the highest control setting, comprising 26 generated motions paired with their neutral references. Action-preservation metrics are evaluated separately using 256 HumanML3D test captions sampled with a fixed random seed.


User-study selection. We selected 11 examples from our available TF, SFT, and SFT+TF qualitative results to cover diverse actions, including walking, dancing, jumping, leaping forward, kicking, running, squatting, crawling, and picking up. These examples were selected for action coverage rather than through a predefined random sampling procedure. Accordingly, the study describes participants’ judgments of these selected examples collectively, rather than an unbiased estimate over all generated motions or a separate evaluation of each approach.


We will clarify these protocols and details in the manuscript. We hope these clarifications address the reviewer's questions and concerns.

AuthorsOpenReview ↗

Response to Reviewer 5bv3 - Weakness 2 and Question 2

We thank the reviewer for suggesting this comparison. We evaluated direct DMP, DMP followed by contact correction (DMP+IK), DMP with inference-time guidance (DMP+TF), and guidance followed by the same contact correction (DMP+TF+IK). At matched achieved exaggeration: (i) TF consistently improves reported action-preservation metrics and reduces motion jerk compared with direct DMP exaggeration, although its effect on foot sliding varies by backbone; and (ii) compared with DMP+IK, TF achieves higher R@3 (better text-motion alignment) and lower jerk across all three backbones, while the remaining metrics show a backbone-dependent trade-off.

We evaluated three pretrained backbones on 40 prompts (26 ExMoCap actions and 14 additional prompts), using two seeds. All methods shared the same neutral motion for each prompt and seed. We generated TF outputs at K={2,3,6,10} and adjusted the direct-DMP and DMP+IK strengths to match achieved root-relative range of motion (ROM).

For correction, we used the damped-Jacobian IK solver from the MoMask codebase. The shared neutral supplied stationary foot-contact intervals and bone lengths. We constructed fixed foot targets during these intervals, blended them into surrounding frames, and adjusted the leg rotations while preserving the root trajectory and upper-body positions. Because IK correction can change ROM, we checked that the maximum ROM-gain difference across all four outputs remained within the original matching tolerance. Table 1 compares corresponding cases within each metric. Δ_phys uses per-action normalization pooled across all four methods.

Table 1. Refinement methods at matched achieved ROM.

BackboneMethodELRA@1 ↑ρ_clf ↑ρ_phys ↑Δ_phys ↑Nov. ↑R@3 ↑FID ↓Div. ↑MMDist ↓Skate ↓Jerk ↓
MDMDMP0.2000.0000.1662.2570.0410.70615.6857.9763.2970.06445.532
DMP+IK0.2000.0000.1662.1830.0250.73115.3187.9923.2100.03046.250
DMP+TF0.2000.0710.1672.1590.0330.73815.5298.0523.2480.07235.362
DMP+TF+IK0.2000.0000.1551.9020.0190.71315.4598.0803.2380.03436.986
HY-Motion-1BDMP0.2000.0260.6032.6330.1160.56716.6746.5974.1250.068192.003
DMP+IK0.2060.0260.6032.6340.1070.57615.9606.7084.0690.058199.838
DMP+TF0.2180.0510.6032.6420.1100.58015.6316.6944.0510.067172.137
DMP+TF+IK0.2240.0510.6012.6420.1040.57115.2986.7794.0230.057181.261
KIMODODMP0.2160.0280.2612.5410.0940.51820.5977.5164.2700.041140.463
DMP+IK0.2160.0280.2622.5460.0800.53618.2117.6184.1230.022148.364
DMP+TF0.2160.0280.2622.5310.0970.55419.4687.6024.1700.032120.194
DMP+TF+IK0.2160.0280.2662.6000.0900.57118.5317.5954.1170.024127.254

This distinction between DMP+TF and DMP+IK also helps explain their complementary strengths: IK targets foot contacts through leg adjustments, whereas TF uses the pretrained motion prior to regularize full-body motion during generation. Therefore, TF’s higher R@3 and lower aggregate jerk provide evidence of benefits beyond contact correction, while IK offers stronger foot stability. This motivates combining the two. Specifically, adding IK after TF further demonstrates their complementarity. Relative to TF alone, DMP+TF+IK reduces foot sliding and improves FID and matching distance across all three backbones. Compared with DMP+IK, the combined approach retains lower aggregate jerk across all backbones and improves several alignment or distributional measures, particularly on HY-Motion.

We thank the reviewer for this helpful question, which allowed us to clarify the contribution of our training-free approach. We hope these comparisons address the reviewer’s concern.

Reviewer_qi1NOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper focuses on a new task in the motion generation field, which aims at controllable exaggeration. The authors introduce both training-time adaptation and test-time adapation. Specifically, during training, the authors first establish a new dataset, ExMocap, which contains 5 different exaggeration level and 26 different actions. To finetune the existing diffusion or flow matching model without losing generalizability. The authors train on both ExMocap and the original training dataset for each model and use LoRA to preserve its original capability. During inference time, the authors proposed a DMP-based method to guide the denoising direction to achieve test-time exaggeration.

Critical strengths
  1. Although the exaggeration can be regarded as a specific type of stylization, the authors provide some theoritical analysis about the difference between natural motion and the motion sequences that used in animation. This direction is important.
  2. The overall writing is satisfying, ensuring that its content is easily understandable for readers.
  3. The authors proposed a comprehensive benchmark for the new proposed task, which contributes a lot to the future research.
Critical weaknesses
  1. The qualitative results are not satisfying. Although the generated videos show higher exaggeration, their naturalness is highly degraded, including frequent self-penetration and extra-articular movement. In addition, the ablation study should also introduce qualitative comparison with more demo videos.
  2. There is few illustration about the dataset quality. For example, the authors should provide more cases, including the same action with different exaggeration level and the same exaggeration level with different action type.
  3. As for the user study, current aspects do not contain naturalness or motion quality. These two dimensions are necessary.
  4. The authors should ablate on the size of SFT dataset.
  5. Since the topic is very similar to motion stylization, the authors are supposed to implement some previous method to justify the effectiveness of the proposed new method.
Questions for authors

Please refer to the weakness part.

7 comments
AuthorsOpenReview ↗

Response to Reviewer qi1N - Weakness 4

We thank the reviewer for this suggestion. We agree that evaluating sensitivity to the SFT dataset size would strengthen the analysis. However, our dataset is already substantially smaller than those used to pretrain the evaluated backbones, while collecting more graded motion-capture performances is costly. During data collection, we incrementally fine-tuned the models as recordings became available. These preliminary experiments showed weaker exaggeration control with smaller subsets, motivating the final dataset size as a practical minimum for effective SFT. We have now clarified this limitation and identified a systematic study of data efficiency as future work in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer qi1N - Weakness 5

We thank the reviewer for suggesting a comparison with prior motion stylization methods. Although motion stylization could be related to our work, it addresses a completely different problem: existing methods typically transfer a reference style or condition generation on predefined style labels. Instead, our work targets the animation principle of exaggeration as a controllable intensity and studies the degree of exaggeration applied to an action, while preserving the underlying action. This differs from settings in which the desired change is specified through a reference motion or a style label.

To our knowledge, no existing motion-stylization method provides a directly comparable exaggeration control under the same setting; implementing one would require introducing a style reference or supervision unavailable to our method, making the comparison difficult to interpret. Therefore, we evaluate against three state-of-the-art motion-generation backbones and compare our approaches with their pretrained baselines and direct DMP exaggeration.

We hope this clarification addresses the reviewer’s concern. We have now clarified this distinction and further strengthened the discussion of motion stylization in the revised manuscript. We would also be happy to consider any specific relevant baseline the reviewer believes supports the same task and control setting.

AuthorsOpenReview ↗

Response to Reviewer qi1N - Weakness 3

We thank the reviewer for highlighting this limitation. We agree that the original user study does not directly assess perceived naturalness or overall motion quality. As described in Sec. 4.2 and Appendix E, it evaluates recognition of exaggeration levels and preservation of the underlying action. The paper separately reports quantitative measures such as FID, diversity, MMDist, and foot-skate (Table 1 in the paper), but these do not replace subjective judgments of naturalness and quality.

To address this concern, we extended the user study by adding explicit questions on both dimensions for our active participant pool. The follow-up questionnaire covers seven videos, each presenting an original motion on the left and its exaggerated counterpart on the right. For each video, participants rated relative naturalness on a three-point scale (much less natural, slightly less natural, and about equally natural) and overall motion quality on a five-point scale (very poor, poor, fair, good, and excellent). Therefore, each participant provided 14 ratings.

The follow-up study included 28 participants: 15 software engineers, 4 machine learning engineers, 3 animators, 3 product owners, 2 modeling artists, and 1 project manager. Across their responses, 69.4% of naturalness ratings selected “about equally natural,” and 67.9% of quality ratings selected “fair,” “good,” or “excellent.” These results indicate that the exaggerated motions were generally perceived as similarly natural to their original counterparts, while overall quality was rated fair or better in approximately two-thirds of evaluations. Table 1 (below) presents the follow-up user study results on motion quality and naturalness.

We will include these additional findings in an extended version of the user-study section, complementing the existing exaggeration and action-preservation results.

We hope this additional evaluation of motion quality and naturalness addresses the reviewer’s concern.


Table 1. Follow-up user-study results on motion quality and naturalness. Percentages are pooled across participant–video ratings, with counts shown in parentheses.

ItemValue
Participants28
Video examples per participant7
Number of naturalness ratings196
Number of motion-quality ratings196
About equally natural69.4% (136/196)
Motion quality: fair or better67.9% (133/196)
Reviewer_Sn3GOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces controllable exaggeration for pretrained text-to-motion models. For this, the authors propose two approaches: i) supervised fine-tuning on a newly collected motion-capture dataset with five exaggeration levels, and ii) training-free guidance toward exaggerated targets obtained by scaling the forcing terms of dynamic movement primitives (DMPs). The two approaches can also be combined.

The methods are evaluated on three motion generators (MDM, HY-Motion, and KIMODO) using exaggeration classifiers, geometric measures, standard motion-generation metrics, and a user study. The results show that the methods can produce progressively exaggerated motions while retaining the underlying action to varying degrees. Supervised fine-tuning generally gives the strongest results in terms of exaggeration control and motion quality.

Critical strengths
  1. I found the qualitative results convincing. The generated animations look good, and the differences between exaggeration levels are clearly visible in several examples. I also like the idea of introducing an additional control to existing motion generators rather than training a new model from scratch. This could be useful for character animation, where realistic motion is not always the desired outcome.

  2. The paper provides a fairly extensive evaluation across three different motion generators, covering both diffusion and flow-matching models. The supervised results are particularly promising. For HY-Motion 1B, ELRA improves from 0.354 to 0.938 while R@3 remains almost unchanged (Table 1). I also appreciate that the authors investigate both supervised and training-free approaches, since these are applicable in different settings.

Critical weaknesses
  1. Motivation: While I liked the qualitative results, I found the practical motivation somewhat underdeveloped. The introduction largely motivates the work through the 12 principles of animation, but it remains unclear where controllable exaggeration would be useful in practice. The examples mainly show increasingly pronounced movements without demonstrating a particular artistic goal. I think a clearer discussion of potential use cases and how this control could benefit animation workflows would help establish the significance of the work.

  2. Related work and experimental comparisons: I think the paper could be better positioned with respect to existing work on motion exaggeration and magnification. For example, Dorkenwald et al. (CVPR 2020) study controllable magnification of human posture differences using a learned generative representation. More directly related to animation, The Cartoon Animation Filter (Wang et al., SIGGRAPH 2006) provides strength-controlled motion exaggeration, while STyMo (Ponton et al., SIGGRAPH 2026) supports posture intensity and temporal exaggeration.

I would not expect all these approaches to be reproduced as baselines, given their different settings. However, comparing direct DMP exaggeration with DMP followed by generative guidance does not establish whether DMP is preferable to simpler exaggeration methods. A comparison using the same generative refinement at comparable exaggeration levels would help clarify the benefit of the proposed construction.

Questions for authors
  1. Could the authors elaborate on the intended use cases for controllable exaggeration? A concrete example where this control helps achieve a particular artistic goal, rather than simply making a movement more pronounced, would strengthen the motivation. It would also be useful to understand what advantages this offers over existing motion-editing approaches.

  2. Have the authors considered comparing DMP-based exaggeration with simpler approaches, such as residual scaling or The Cartoon Animation Filter, under the same generative guidance? A comparison at similar perceived exaggeration levels would help clarify whether the DMP formulation provides an advantage in motion quality and action preservation.

  3. How were the five exaggeration levels defined for the performers, and was their ordering independently validated? How is the DMP scale KK mapped to these levels? Since exaggeration can be expressed differently across actions and performers, it would be useful to understand how well the proposed metrics reflect perceived exaggeration.

  4. How were the original motion recordings split between classifier training, validation, and evaluation? Were these splits made before windowing and augmentation? Furthermore, the user study mainly evaluates whether the original action remains recognizable. Does this sufficiently capture motion-intent preservation, given that an action can remain recognizable while its manner or expressive characteristics change?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Sn3G - Question 4

We thank the reviewer for this careful question.

1- The answer to the first part of question 4 - Classifier splits

The split in our data preparation pipeline is made after windowing and augmentation. Specifically, we first segment the ExMoCap recordings into windows and generate the mirrored and jittered variants. Then, we split the data by window, keeping each window and its augmented variants together in either the training partition or the validation partition. The classifiers are trained on the training partition. For the final classifier evaluation, we divide the validation pool into two clip-level subsets divided by exaggeration level. Therefore, different windows from the same original recording can appear in the training and validation partitions, and variants of a validation window can appear in both classifier subsets.

2- The answer to the second part of question 4 - Action recognition versus motion-intent preservation

An action can remain recognizable even when its manner, expressive characteristics, or perceived intent changes. In our setting, exaggeration is specifically designed to modify manner and expressivity while preserving its underlying action type. To assess this, participants viewed a source motion alongside its exaggerated counterpart and rated how well the counterpart preserved the source action type on a four-point scale from “very well preserved” to “not preserved.” Therefore, the study measures whether the semantic action remains consistent as its expressive manner changes. We have now clarified this distinction in the manuscript and replace the ambiguous term “motion-intent preservation” with the more precise “action-type preservation.”

AuthorsOpenReview ↗

Response to Reviewer Sn3G - Weakness 2 (Part 1)

We thank the reviewer for highlighting these related works. Here, we discuss and clarify their relationship to our setting.

Dorkenwald et al. (CVPR 2020) study a different task which involves magnifying posture differences between subjects in images. Their method operates in image space and disentangles appearance from posture to visualize deviations between observed poses. However, our task requires generating and controlling exaggeration over a temporal sequence of full-body 3D motion while preserving the underlying action and motion quality. We believe this is a completely different task. These differences in input, output, and objective prevent a direct comparison. In fact, adapting their method would require extending an image-based posture representation to temporally consistent 3D motion generation. Therefore, we consider this work conceptually related but not a directly compatible baseline. We will clarify this distinction in the related-work discussion.

STyMo also addresses a different task which is transferring a reference motion style using paired neutral and stylized examples. It learns a style-specific transformation, and its posture-intensity and temporal-exaggeration controls adjust the strength of the learned static and dynamic style components. However, our work studies exaggeration as a controllable animation principle within pretrained text-to-motion generators. In particular, our training-free method derives an action-specific exaggeration target directly from the generator’s own neutral motion, without requiring paired style examples or training a style-specific model. Therefore, applying STyMo would require paired data and additional training for the desired style, introducing different data and model requirements from our setting. Although STyMo provides relevant intensity controls, these controls operate within its learned style transformation not the exaggeration motion which is the target of our study. We will clarify this distinction in the related-work discussion.

However, we agree that the Cartoon Animation Filter is more directly applicable as an alternative target within our guidance framework. Therefore, in our training-free approach, we replaced the DMP target with CAF and evaluated CAF+TF against DMP+TF on MDM, HY-Motion-1B, and KIMODO. The details of these comparisons are presented below in our next response which jointly addresses the second part of Weakness 2 and Question 2, titled “Response to Reviewer Sn3G - Weakness 2 (Part 2) and Question 2.”

AuthorsOpenReview ↗

Response to Reviewer Sn3G - Weakness 2 (Part 2) and Question 2.

We thank the reviewer for suggesting this comparison. We implemented the Cartoon Animation Filter (CAF) [1] as an alternative guidance target and evaluated CAF+TF (training-free approach with CAF target) against DMP+TF (training-free approach with DMP target) on MDM, HY-Motion-1B, and KIMODO. We adapted CAF to the same pose channels used by DMP. Both methods used identical prompts, seeds, neutral motions, pretrained generators, initial noise, and guidance settings.

We report two complementary comparisons: a five-setting strength sweep (Table 1) and a matched-intensity comparison (Table 2). Because CAF and DMP strength parameters have different meanings, equal numerical settings do not imply comparable exaggeration. Therefore, matching achieved physical amplification (Table 2) provides a fairer comparison of quality under otherwise identical guidance conditions.

First, we tested five increasing strength settings per method on 26 actions across all three backbones (Table 1). CAF+TF produced larger physical motion ranges, more consistent increases in physical intensity, and greater feature-space deviation from the neutral motion (Novelty). At the highest tested setting per method, DMP+TF achieved better text-motion alignment, FID, and diversity across all three backbones. These results describe each method’s exaggeration-quality trade-off and motivate a comparison at matched intensity.

Second, we matched DMP+TF and CAF+TF outputs from the same action prompt, seed, and neutral reference by their root-relative range-of-motion (ROM) gain over that reference (Table 2). A small experiment identified low, medium, and high gains achievable by both methods. This yielded 35 matched pairs for MDM, 42 for HY-Motion-1B, and 54 for KIMODO across the three intensity ranges. Novelty and foot skate used all matched pairs. R@3, FID, diversity, and MMDist used the same 32 selected paired cases per backbone, ensuring evaluation on corresponding motions.

At matched intensity (Table 2), CAF+TF produced greater Novelty across all three backbones. DMP+TF generally achieved better action alignment on HY-Motion-1B and KIMODO, while MDM showed a mixed trade-off. DMP+TF produced less foot sliding on MDM, CAF+TF produced less on KIMODO, and both performed similarly on HY-Motion-1B. Therefore, The target affects the balance between deviation from the neutral motion, action alignment, and foot sliding, with the outcome depending on the backbone.

Overall, DMP+TF offers a competitive balance between physical amplification (exaggeration), action alignment, and motion quality, while neither target consistently outperforms the other across backbones and metrics. We have added both tables and clarified this conclusion in the manuscript. We hope these comparisons address the reviewer’s concern.

Table 1. DMP+TF versus CAF+TF across five strength settings.

ModelMethodELRA@1 ↑ρ_clf ↑ρ_phys ↑Δ_phys ↑Nov. ↑R@3 ↑FID ↓Div. ↑MMDist ↓Skate ↓
MDMDMP+TF0.2030.0620.2300.5310.0300.71915.9597.8383.3020.094
MDMCAF+TF0.2000.0760.4471.1300.3420.23437.3584.8336.9170.150
HY-Motion-1BDMP+TF0.2130.0290.4491.0350.0890.41418.7536.3074.8900.057
HY-Motion-1BCAF+TF0.2810.2390.8642.2080.4560.13382.4401.9668.8150.064
KIMODODMP+TF0.2030.0210.2620.7650.0750.56324.5517.6864.6150.036
KIMODOCAF+TF0.131-0.2320.5151.2000.3550.15644.2104.1497.2390.038

Table 2. DMP+TF versus CAF+TF at matched physical intensity.

ModelMethodNov. ↑R@3 ↑FID ↓Div. ↑MMDist ↓Skate ↓
MDMDMP+TF0.0180.65612.7917.7763.0440.114
MDMCAF+TF0.0580.75012.1017.6743.4320.189
HY-Motion-1BDMP+TF0.0540.56316.1426.1634.1320.0660
HY-Motion-1BCAF+TF0.0580.53117.3776.2154.2940.0671
KIMODODMP+TF0.0490.53118.0277.4124.1090.043
KIMODOCAF+TF0.0960.40620.6396.8694.7490.035

Local Reference:

[1] Wang, Jue, et al. “The Cartoon Animation Filter.” ACM Transactions on Graphics 25.3 (2006), 1169–1173.

Reviewer_J5dXOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies controllable exaggeration in human motion generation through supervised fine-tuning and training-free guidance. It introduces ExMoCap, a motion-capture dataset covering 26 actions at five exaggeration levels, and a DMP-based approach that guides pretrained generators toward exaggerated motion targets. Experiments evaluate exaggeration control, action preservation, and motion quality, showing that fine-tuning improves level control, while combining it with training-free guidance generally yields stronger physical exaggeration at the expense of action preservation or motion quality.

Critical strengths
  1. The paper studies a relevant problem that is not well captured by standard text-to-motion evaluation. Treating exaggeration as an explicit animation principle provides a useful perspective on expressive motion generation.
  2. The paper evaluates both training-time and training-free approaches on three backbones, including diffusion and flow-matching models. It also includes ablations, standard motion-generation metrics, exaggeration-specific metrics, qualitative results, and a user study.
Critical weaknesses
  1. The manuscript motivates exaggeration through animation principles and discussions with professional animators, its practical benefits for animation workflows remain insufficiently demonstrated.
  2. The generalization of the SFT approach is not sufficiently clear. ExMoCap contains 26 actions performed by two actors, and I could not determine from the paper whether the main SFT evaluation uses actions held out from fine-tuning. Evaluation on clearly held-out actions would provide stronger evidence that the model has learned transferable exaggeration behavior rather than mainly adapting to the ExMoCap action set.
  3. Part of the exaggeration evaluation is closely tied to ExMoCap. The classifiers used for ELRA and rho_clf are trained on ExMoCap, which is also the dataset used to supervise SFT. The paper states that these classifiers are frozen during evaluation, but the relationship between their training/validation data and the SFT data is not sufficiently clear to me.
  4. The user study has limited coverage: exaggeration control is assessed using five video examples, while action preservation is assessed using six examples, with an uneven distribution across backbones. Repeated ratings of these examples across 20 participants do not establish generalization across actions and models.
Questions for authors
  1. Were the SFT models evaluated on actions that were completely held out from ExMoCap fine-tuning? If so, please clarify the action split and indicate which results correspond to unseen actions. If not, how should the current results be interpreted with respect to generalization beyond the ExMoCap action set?
  2. How were the ExMoCap samples split when training the exaggeration-level classifiers? In particular, are the classifier training and validation samples disjoint from the SFT data in terms of motion clips, action instances, or actors?
  3. SFT is controlled using five text-based intensity levels, whereas DMP and TF are controlled using the DMP scale K. How were these different control scales chosen or aligned for the quantitative comparisons?
  4. Table 1 shows substantial FID increases after adding TF to both the baseline and SFT models across all three backbones. Although exaggeration can legitimately shift motions away from the reference distribution, TF also worsens MMDist and foot-skate scores in these comparisons and substantially reduces R@3 for HY-Motion and KIMODO. Could the authors clarify whether the additional exaggeration justifies these losses and whether they persist at matched exaggeration intensity? The claim that the methods remain competitive in action preservation and motion quality should be qualified accordingly.
7 comments
AuthorsOpenReview ↗

Response to Reviewer J5dX - Weakness 4

We thank the reviewer for raising this point. At submission, we had collected responses from 20 participants. Since then, we have collected responses from 5 additional general participants and 34 industry professionals, including animators, artists, product owners, and software engineers. This brings the total to 59 participants, with 25 in the general group and 34 in the expert group. In the expert group, participants selected the highest requested intensity as the most exaggerated in 78.2% of trials and rated the underlying action as mostly or very well preserved in 86.3% of trials. The corresponding rates in the general group were 82.3% and 78.7%. These results indicate that participants in both groups generally perceived the intended exaggeration and recognized the original action in the motions they evaluated. Tables 1 and 2 below report the results for the general and expert groups, respectively. The study included 16 evaluation items per participant. Across 59 participants, this corresponds to 944 possible participant–item ratings, with 295 for exaggeration control (5 items), 295 for neutral motion detection (5 items), and 354 for action preservation (6 items). The additional participants provide more judgments on the same set of examples.

Moreover, based on related work in character motion generation, we believe the number of participants and evaluation items is sufficient for the intended scope of our study. For example, Wu et al. (2025) [1] conducted a perceptual study of stylized motion generation with 16 participants and 10 text/style inputs. Goel et al. (2024) [2] evaluated motion editing with 19 participants and 9 edited-motion comparisons per baseline. These are only two examples from motion generation, editing, and stylization that put the number of participants and examples in our study into context. Other focused perceptual evaluations [3, 4, 5] used sample counts on the same order of magnitude. Overall, our study includes more samples than most of these studies, which supports our view that the sample size is sufficient for a focused perceptual evaluation.

Also, please note that we designed the user preference study as a focused, 5–10-minute perceptual evaluation to complement the broader quantitative experiments. As the reviewer notes, the motion examples are unevenly distributed across backbones. However, our study asked whether our approach could exaggerate a motion while preserving its original action, regardless of which backbone generated the source motion. The results address this question for the examples shown in the user study questionnaire. We hope the additional results from the general and expert participants help address the reviewer’s concern.

We have clarified this scope and added the expert results to the revised manuscript.

Table 1. General user-study results for controllable exaggeration and action preservation (25 participants).

BackboneExag. actionsHighest intensityPreserv. actionsAction preserved
KIMODO386.7% (65/75)583.2% (104/125)
HY-Motion179.2% (19/24)156.0% (14/25)
MDM172.0% (18/25)0—
Overall582.3% (102/124)678.7% (118/150)

Table 2. Expert user-study results for controllable exaggeration and action preservation (34 participants).

BackboneExag. actionsHighest intensityPreserv. actionsAction preserved
KIMODO379.4% (81/102)587.6% (149/170)
HY-Motion182.4% (28/34)179.4% (27/34)
MDM170.6% (24/34)0—
Overall578.2% (133/170)686.3% (176/204)

Highest intensity is the percentage of valid responses that identified the highest requested intensity as the most exaggerated. Action preserved is the percentage of valid ratings marked very well preserved or mostly preserved. Counts are shown in parentheses. A dash indicates that no example was evaluated.


Local References:

[1] Wu, Linjun, et al. "Semantically Consistent Text-to-Motion with Unsupervised Styles." Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers. 2025.

[2] Goel, Purvi, et al. "Iterative motion editing with natural language." ACM SIGGRAPH 2024 Conference Papers. 2024.

[3] Sawdayee, Haim, et al. "Dance Like a Chicken: Low‐Rank Stylization for Human Motion Diffusion." Computer Graphics Forum. 2026.

[4] Lee, Eunjong, et al. "Controllable Long-term Motion Generation with Extended Joint Targets." 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2026.

[5] Ponton, Jose Luis, et al. "STyMo: Fast and Controllable Few-Shot Motion Style Transfer." ACM Transactions on Graphics (TOG) 45.4 (2026): 1-13.

AuthorsOpenReview ↗

Response to Reviewer J5dX - Weakness 3 and Question 2

We thank the reviewer for requesting this clarification. The exaggeration-level classifiers are trained separately from the motion generators with completely different neural architectures as explained in Appendix B.1 in the manuscript, but their training data are not disjoint from the SFT training data. For each backbone (MDM, HY-Motion, and KIMODO), both the classifier and the motion generator use the training partition of the corresponding ExMoCap representation. The classifiers are not used in the SFT training objective or inference-time guidance and, as mentioned, they are frozen evaluators applied to newly generated motions.

The dataset partition is constructed at the source-window level. Specifically, the split in our data preparation pipeline is made after windowing and augmentation. We first segment the ExMoCap recordings into windows and generate the mirrored and jittered variants. Then, we split the data by window, keeping each window and its augmented variants together in either the training partition or the validation partition. The classifiers are trained on the training partition. For the final classifier evaluation, we divide the validation pool into two clip-level subsets divided by exaggeration level. Therefore, different windows from the same original recording can appear in the training and validation partitions, and variants of a validation window can appear in both classifier evaluation (test and validation) subsets.

To place our approach in context, the paper "Dance Like a Chicken" [1] also uses a separately trained recognition classifier, training its style evaluator on 100STYLE, the same dataset used for generative adaptation. Nevertheless, we agree that freezing the classifier alone does not establish data independence. We will clarify the shared training data and the scope of the splits in the manuscript. Accordingly, ELRA and ρclf\rho_{clf} measure recognition of the requested exaggeration level using an ExMoCap-trained evaluator; they do not independently establish generalization to unseen recordings, action instances, or actors. Therefore, we complement these scores with classifier-free physical-intensity metrics and human evaluation (Table 1 in the manuscript).

We hope this clarification addresses the reviewer’s concern.


Local Reference:

[1] Sawdayee, Haim, et al. "Dance Like a Chicken: Low‐Rank Stylization for Human Motion Diffusion." Computer Graphics Forum. 2026.

AuthorsOpenReview ↗

Response to Reviewer J5dX - Question 3

We thank the reviewer for this question. For the exaggeration-control evaluation, we pair the controls by their order. Baseline and SFT use five text-based intensity levels: neutral, slightly exaggerated, moderately exaggerated, very exaggerated, and extremely exaggerated. Direct DMP and TF use the action caption without an intensity descriptor and sweep K = 1, 2, 3, 4.5, and 6. For SFT+TF, each text level is paired with the K value at the corresponding position in this ladder.

This protocol gives each approach five ordered settings for evaluating its response to increasing control strength. However, since SFT and TF can be used independently or combined, we evaluate how each setting responds to its corresponding control or combination of controls. This means we did not calibrate a text level and its paired K value to produce equal physical or perceived exaggeration. Therefore, these comparisons assess controllability along each method’s own control axis, rather than performance at matched achieved intensity. We will make this distinction and the exact control settings explicit in the manuscript.

The table below summarizes the correspondence between the five text-based intensity levels and the DMP scale values used in the evaluation.


Table 1. Ordered control settings for each method. Corresponding settings are not calibrated to equal achieved exaggeration.

LevelBaseline/SFT text descriptorDMP/TF scale KSFT+TF
0In a neutral, natural way1Text level 0 + K=1
1Slightly exaggerated2Text level 1 + K=2
2Moderately exaggerated3Text level 2 + K=3
3Very exaggerated4.5Text level 3 + K=4.5
4Extremely exaggerated6Text level 4 + K=6

Comments on the paper

3 comments
AuthorsOpenReview ↗

Thank you to the reviewers and our plan for responses

We thank all reviewers for taking the time to read our paper carefully and provide helpful feedback. We are working through your comments point by point and will post each response as it is ready, so you have time to consider our replies and let us know if any concerns remain.

Area_Chair_6NdzOpenReview ↗

Please engage in the discussion

Dear Reviewers,

Thank you for your reviews and contributions so far. The authors have now posted their response, and I would like to kindly remind you to review it and engage in the discussion where appropriate.

Please also consider whether the authors' clarifications address any of your concerns and update your assessment if warranted.

Thank you for your time and effort.

Best regards,

Your AC.

AuthorsOpenReview ↗

Final message to reviewers with updated manuscript references

We thank all four reviewers for their careful feedback and constructive suggestions. In our response, we conducted additional quantitative experiments, expanded the perceptual evaluation, added qualitative comparisons, and clarified the scope of our contributions and claims. We summarize these additions below and point to the relevant locations in the revised manuscript.

  • Contribution, related work, and practical motivation - Reviewers J5dX, Sn3G, qi1N, and 5bv3. In our responses, we clarified how our formulation differs from reference-style transfer, motion stylization, and existing editing frameworks, including LoRA-MDM, STyMo, GMD, and Deep Motion Warping. Our contribution is the formulation and evaluation of controllable exaggeration within pretrained motion generators. We also clarified the intended application of helping animators explore expressive variations of an action through text or continuous inference-time control. Production integration and artist-centered workflow evaluation remain future work. The manuscript provides the practical motivation in Sec. 1, pp. 1-2, the related-work context in Sec. 2, pp. 2-3, and the training-time and inference-time formulations in Secs. 3.2-3.3, pp. 4-7.

  • Generalization beyond ExMoCap actions - Reviewer J5dX. We evaluated 10 actions outside ExMoCap fine-tuning, using five requested intensity levels and three seeds, yielding 150 motions per backbone and method. SFT improved all five reported exaggeration metrics across the three backbones, with stronger transfer for HY-Motion-1B and KIMODO and more limited transfer for MDM. We added qualitative examples in the supplementary material and discuss the shared rowing failure case in the manuscript. See Appendix E, pp. 25–26, Table 10 (p. 26).

  • Comparisons at matched physical exaggeration intensity - Reviewers Sn3G and 5bv3. We conducted two complementary comparisons to separate the effects of the exaggeration target and generative guidance. First, we compared DMP+TF with CAF+TF (Cartoon Animation Filter) under identical guidance, using strength sweeps and range-of-motion (ROM) matching. This tests how the choice of target affects the result at comparable spatial amplification. See Appendix E, pp. 26–27, Tables 11 and 12 (p. 27). Second, we compared direct DMP, DMP+IK (contact correction), DMP+TF, and DMP+TF+IK on 40 action prompts across four intensities, matching achieved ROM. This tests whether guidance offers benefits beyond weaker direct editing or conventional contact correction. See Appendix E, pp. 27–28, Table 13 (p. 28).

  • Motion artifacts and exaggeration-quality trade-offs - Reviewers J5dX, qi1N, and 5bv3. We clarified that stronger exaggeration can reduce action alignment and introduce artifacts. For the walking and dancing examples, we added skeleton-based correction targeting backward knee bending and body-segment intersections. Measurements and videos before and after correction show reductions in the checked violations while retaining most of the original ROM. See Sec. 4.2, pp. 8–9, and Table 1 (p. 8) for the quantitative trade-offs and Appendix F, pp. 28–29, Tables 16 (p. 29) and 17 (p. 30) for the correction and its measured effects.

  • Perceptual evaluation and preservation terminology - Reviewers J5dX, qi1N, Sn3G, and 5bv3. The expanded study includes 59 participants, comprising 34 industry professionals and 25 general participants. The protocol is described in Appendix E, p. 25, with group results in Tables 14 (p. 28) and 15 (p. 29) and representative questionnaire items in Fig. 12 (p. 43). A follow-up with 28 participants assessed naturalness and overall motion quality across seven videos. Of the naturalness ratings, 69.4% selected “about equally natural,” and 67.9% of the quality ratings selected “fair” or better. See Appendix E, p. 25, and Table 9 (p. 26).

  • Qualitative ablations and dataset illustrations - Reviewers qi1N, 5bv3, and Sn3G. We added six supplementary videos showing component effects, TF and text controllability, DMP regularization, multi-keyframe reconstruction, and DMP+TF versus CAF+TF. An additional supplementary video presents nine recorded actions across exaggeration levels, supporting comparisons both within an action and across actions. Related qualitative results are discussed in Appendix E, p. 24, and shown in Figs. 6–11 (pp. 37–42). The ExMoCap recording protocol and analysis of the recorded levels are described in Appendix D.4, pp. 21–22, Tables 6 (p. 22) and 7 (p. 23).

We hope these additional experiments, expanded evaluations, supplementary videos, and methodological clarifications help the reviewers assess both the capabilities and current limitations of our approach.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.