- 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
-
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.
-
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
-
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.
-
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 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
-
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?
-
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 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.
-
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
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
| Partition | Original windows | Mirrored copies | Jittered copies | Total clips | Original recordings |
|---|---|---|---|---|---|
| Training | 1874 | 1874 | 1874 | 5622 | 132 |
| Validation | 208 | 208 | 208 | 624 | 109 |
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 detail | Count |
|---|---|
| Early-stopping (validation) clips | 310 |
| Unique recordings in validation set | 101 |
| Test clips | 314 |
| Unique recordings in test set | 101 |
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.
| Classifier | Test accuracy |
|---|---|
| MDM | 97.44% |
| HY-Motion | 98.73% |
| KIMODO | 98.73% |
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.
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.
| Backbone | Method | ELRA@1 ↑ | ρ_clf ↑ | ρ_phys ↑ | Δ_phys ↑ | Nov. ↑ | R@3 ↑ | FID ↓ | Div. ↑ | MMDist ↓ | Skate ↓ | Jerk ↓ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MDM | DMP | 0.200 | 0.000 | 0.166 | 2.257 | 0.041 | 0.706 | 15.685 | 7.976 | 3.297 | 0.064 | 45.532 |
| DMP+IK | 0.200 | 0.000 | 0.166 | 2.183 | 0.025 | 0.731 | 15.318 | 7.992 | 3.210 | 0.030 | 46.250 | |
| DMP+TF | 0.200 | 0.071 | 0.167 | 2.159 | 0.033 | 0.738 | 15.529 | 8.052 | 3.248 | 0.072 | 35.362 | |
| DMP+TF+IK | 0.200 | 0.000 | 0.155 | 1.902 | 0.019 | 0.713 | 15.459 | 8.080 | 3.238 | 0.034 | 36.986 | |
| HY-Motion-1B | DMP | 0.200 | 0.026 | 0.603 | 2.633 | 0.116 | 0.567 | 16.674 | 6.597 | 4.125 | 0.068 | 192.003 |
| DMP+IK | 0.206 | 0.026 | 0.603 | 2.634 | 0.107 | 0.576 | 15.960 | 6.708 | 4.069 | 0.058 | 199.838 | |
| DMP+TF | 0.218 | 0.051 | 0.603 | 2.642 | 0.110 | 0.580 | 15.631 | 6.694 | 4.051 | 0.067 | 172.137 | |
| DMP+TF+IK | 0.224 | 0.051 | 0.601 | 2.642 | 0.104 | 0.571 | 15.298 | 6.779 | 4.023 | 0.057 | 181.261 | |
| KIMODO | DMP | 0.216 | 0.028 | 0.261 | 2.541 | 0.094 | 0.518 | 20.597 | 7.516 | 4.270 | 0.041 | 140.463 |
| DMP+IK | 0.216 | 0.028 | 0.262 | 2.546 | 0.080 | 0.536 | 18.211 | 7.618 | 4.123 | 0.022 | 148.364 | |
| DMP+TF | 0.216 | 0.028 | 0.262 | 2.531 | 0.097 | 0.554 | 19.468 | 7.602 | 4.170 | 0.032 | 120.194 | |
| DMP+TF+IK | 0.216 | 0.028 | 0.266 | 2.600 | 0.090 | 0.571 | 18.531 | 7.595 | 4.117 | 0.024 | 127.254 |
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.
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.