Source-linked AI summary
Motion Style Slider: Endpoint-Supervised Continuous Style Control for Human Motion Diffusion
Chen-Chieh Liao, Yichen Peng, Yiyi Cai, Yûi Ono, Hiroki Hanaoka, Erwin Wu, Hideki Koike, Shuichi Kurabayashi
TL;DR
Fine-grained style intensity control remains underexplored, despite the practical need for predictable relative adjustments and evaluation beyond seen style strength. Motion Style Slider uses endpoint-derived directions in a learned motion embedding space to condition diffusion with scalar intensity and latent regularization. Across multiple datasets it improves control-oriented behavior while remaining competitive on realism and content metrics, and evaluates extrapolation against real stronger-style motions.
Problem
Existing methods often treat style as categorical or weakly interpolated, leaving fine-grained intensity control and validation beyond seen style strength underexplored.
Source: §1 · pp. 1–2Method
Motion Style Slider derives a style direction from endpoint motions and conditions diffusion with scalar intensity, using latent regularization for continuous control without intermediate-intensity supervision.
Source: §1 · p. 3, §3.4 · p. 8, §6 · p. 15Results
Across PerMo, Bandai-Namco, and Xia, the method improves style fidelity and slider consistency while remaining competitive on realism and content metrics; an over-reaction benchmark measures performance against real stronger-style motions.
Source: §6 · p. 15Takeaways & Limitations
The over-reaction benchmark evaluates out-of-range performance directly against real stronger-style motions rather than relying only on in-range transfer quality.
Source: §6 · p. 15, §5 · p. 15Training uses only neutral/content and stylized endpoint pairs, while intermediate-intensity motions are typically unavailable.
Source: §3.1 · p. 6
Abstract
from arXiv · showhide
Existing human motion diffusion methods provide strong motion generation quality, and recent style transfer models can inject target style cues, but fine-grained continuous control of style intensity remains underexplored. In production, style intensity is subjective across artists and directors, so the practical requirement is not a universal absolute unit, but a reliable monotonic control axis. We propose Motion Style Slider, a motion-to-motion style transfer framework for endpoint-supervised continuous control. Given a content motion and a style motion, we construct a style direction in a learned motion-style embedding space and condition diffusion generation with a scalar intensity. The training objective combines diffusion denoising with latent intensity regularization to encourage smooth and monotonic style scaling without requiring intermediate-intensity ground-truth motions. Our framework is compatible with pretrained motion diffusion backbones and supports heterogeneous style datasets, including the multi-actor style motion dataset. To test out-of-range usability, we additionally introduce a small real-capture over-reaction extension and evaluate large-intensity behavior against these unseen targets. Experiments measure controllability, interpolation/extrapolation behavior, content preservation, and motion realism, with ablations on direction construction and loss design.
1 Introduction
The introduction identifies reliable, fine-grained style intensity control and real-target extrapolation as open challenges, especially when training data provides only neutral and fully stylized endpoints. Motion Style Slider represents style change as a direction in a learned embedding space and uses scalar intensity conditioning with latent regularization.
1 Introduction
- Continuous style strength is underexplored: existing pipelines often use discrete labels or weak interpolation, and real-motion extrapolation is rarely validated.The practical target is predictable, smooth relative control rather than a universal intensity unit.Source: §1 · p. 2
- The endpoint-only setting pairs neutral-content and fully stylized motions, leaving intermediate-intensity ground truth unavailable for learning smooth, monotonic, extrapolatable variation.This missing supervision is the central learning challenge identified by the introduction.Source: §1 · p. 2
- Style change is modeled as a direction in a learned style embedding space, with scalar α conditioning diffusion alongside content motion and latent intensity regularization.The framework is implemented on an off-the-shelf motion diffusion model with style representations derived from motion-language alignment models.Source: §1 · p. 3
- Evaluation centers on style monotonicity, interpolation and modest out-of-range behavior, content preservation, and motion realism.The introduction frames reliable relative control as the goal, rather than one-shot transfer quality.Source: §1 · p. 3
- The work formulates endpoint-supervised diffusion style transfer as scalar-slider control, and proposes direction conditioning and latent regularization for smooth, monotonic scaling without intermediate-intensity ground truth.Its contributions also include a multi-dataset evaluation protocol, a real-capture over-reaction split, and ablations of direction design and loss terms.Source: §1 · p. 3
2 Related Work
Related work spans text-conditioned motion generation, motion style transfer, and semantic motion representations, but calibrated continuous style intensity remains relatively unexplored. The paper uses a motion-language embedding space to define endpoint style directions and evaluates continuous scaling across datasets and real-capture over-reaction targets.
2.1 Text-to-Motion Generation
- Text-to-motion systems have advanced through tokenized, autoregressive, diffusion, and controllable approaches, but primarily control action semantics rather than calibrated stylistic variation.The cited diffusion approaches improve realism and diversity, while controllable variants add spatial constraints or real-time control.Source: §2.1 · pp. 3–4
2.2 Motion Style Transfer
- Motion style transfer evolved from optimization and learning-based methods to weakly supervised, cross-content, constrained, real-time, and diffusion approaches, yet continuous calibrated intensity remains largely unexplored.Most methods use categorical style conditions or transfer toward a single target instance, evaluating stylization at a fixed operating point.Source: §2.2 · p. 4
2.3 Motion Representation Learning and Style Conditioning
- Motion-language embedding spaces offer semantic directions for control; the paper uses a frozen TMR-based encoder to define endpoint style directions, not to introduce a new representation model.Evaluation spans multiple style datasets and a real-capture over-reaction split for extrapolation beyond observed endpoints.Source: §2.3 · pp. 4–5, Figure 2 · p. 5
3 Method
The method learns continuous relative style control from endpoint pairs, moving through an embedding-space style direction while separately conditioning on content. Latent regularizers shape style trajectories without requiring intermediate-intensity ground-truth motions.
3.1 Problem Formulation
- Endpoint pairs anchor the slider at neutral content for α=0 and the stylized motion for α=1, while α represents relative rather than absolute style strength.Intermediate-intensity motions are typically unavailable.Source: §3.1 · p. 6
3.2 Style Direction in Latent Space
- The method derives a style direction from content and stylized embeddings, using instance, prototype, or hybrid directions to define the continuous style axis.The hybrid direction balances global and instance-specific directions with β.Source: §3.2 · p. 6
3.3 Condition Construction
- The denoiser receives separate content and style conditions, with the style embedding moved along its direction by α.Endpoint-aware intensity sampling increases endpoint supervision while also exposing the model to intermediate values.Source: §3.3 · p. 7
3.4 Training Objective
- Diffusion denoising preserves endpoint realism, while latent linearity and margin-based monotonicity losses shape style scaling, including at α>1.Intermediate samples are not treated as hard motion-space midpoint targets because real midpoint motions are unavailable.Source: §3.4 · p. 8
3.5 Training and Inference
- At inference, users select α to construct a style condition and sample motions, sweeping from interpolation values to extrapolation above 1.The main configuration uses adaptor-only finetuning with motion-conditioned content input and the text branch disabled.Source: §3.5 · pp. 8–9
4 Experiments
Experiments use three benchmark style-motion datasets and a separate real-capture over-reaction set for extrapolation evaluation. Methods share the pair protocol, intensity values, and evaluator.
4.1 Datasets
- The benchmark evaluation uses PerMo, Bandai-Namco, and Xia, with endpoint pairs formed from neutral/content and stylized motions.Data are converted to a shared feature representation.Source: §4.1 · p. 9
- A test-only professional-actor over-reaction set provides stronger real captures matched to content-style identities as unseen extrapolation targets.Only endpoint-consistent valid triplets are retained after quality-control filtering.Source: §4.1 · p. 9
4.2 Implementation Details
- All methods use canonical test pairs and the shared intensity set α∈{0, 0.5, 1.0, 1.5, 2.0}, with metrics computed by a frozen TMR style-encoder head.Ablations may use subsets of the intensity set.Source: §4.2 · p. 9, §4.1 · p. 9
- Baselines use the same pair splits, intensity values, and post-hoc evaluator, keeping metric extraction and prototype construction consistent across methods.Each baseline uses its own trained checkpoint.Source: §4.2 · p. 10
4.3 Baselines and Variants
The evaluation compares endpoint-paired transfer baselines and tests alternative direction modes and regularizers. Metrics cover style control, content preservation, realism, and out-of-range behavior.
4.3 Baselines and Variants
- DeepMotionEditing and MCM-LDM are evaluated on harmonized endpoint pairs, alpha values, and feature-space metrics.MCM-LDM is a latent-diffusion style-transfer baseline without an explicitly optimized endpoint-anchored style direction or scalar intensity objective.Source: §4.3 · p. 10
- Ablations vary prototype, instance, and hybrid directions and test removing Llin or Lmono.These variants isolate direction construction and objective components.Source: §4.3 · p. 10
4.4 Metrics
- Control quality is evaluated over α ∈ {0, 0.5, 1.0, 1.5, 2.0} using style-strength trajectories, linearity error, and monotonicity measures.MonoViol is the averaged positive decrease in style strength across consecutive alpha values.Source: Controllability metrics · pp. 10–11, §4.4 · p. 10
- FMD, CRA, and SRA assess realism, content preservation, and style fidelity, while ExtraErr compares generated α = 2.0 motions with held-out stronger-style targets.Table 2 averages FMD, CRA, SRA, MonoViol, and LinErr over PerMo, Bandai-Namco, and Xia; ExtraErr uses the over-reaction set.Source: §4.4 · p. 10, Standard transfer metrics · p. 11, Extrapolation metric · p. 11
4.5 Main Quantitative Results
Across the three main style datasets and held-out over-reaction targets, the method performs strongly on style-control metrics, while MCM-LDM retains an advantage in realism and content measures.
4.5 Main Quantitative Results
- 0.295 SRA, 0.05 MonoViol, 0.280 LinErr, and 1.109 ExtraErr are the method’s best control-related results in the global comparison.MCM-LDM remains strong on FMD and CRA, indicating a realism/content advantage under this setting.Source: §4.5 · p. 11
- On PerMo, the method improves SRA to 0.140 from MCM-LDM’s 0.080 and lowers MonoViol to 0.03 from 0.22.These are the PerMo values reported for the two methods.Source: §4.5 · p. 11
- On Bandai-Namco and Xia, the method has lower LinErr; its MonoViol is lower on Bandai-Namco but slightly higher on Xia.The reported controllability gains therefore vary with dataset style geometry.Source: §4.5 · p. 11
4.6 Extrapolation to 2× Style Targets
The method is evaluated for extrapolation by comparing generated 2.0-intensity motions with real, stronger-style captures held out from training.
4.6 Extrapolation to 2× Style Targets
- On the over-reaction dataset, ExtraErr is 1.109 for ours, 1.196 for DeepMotionEditing, and 1.174 for MCM-LDM.The generated α = 2.0 motions are compared with real stronger-style targets.Source: §4.6 · p. 11
- The over-reaction clips are test-only matched triplets, with a real stronger capture of the same content-style identity as the target.Actors were instructed to perform over-reactions, and the valid triplets were quality-controlled.Source: §4.6 · p. 11
- The method is closest to the real stronger-style targets among the three compared methods on this dataset.This conclusion follows the reported ExtraErr values.Source: §4.6 · p. 11
4.7 Ablation Studies
Ablations examine direction choices and regularization, while qualitative examples show intensity separation and motion-identity preservation. The broader reference-based example is presented as illustrative rather than a benchmark claim of general transfer.
4.7 Ablation Studies
- On PerMo, instance direction gives the best FMD, SRA, and LinErr, while hybrid direction remains competitive on CRA and MonoViol.Removing either Llin or Lmono weakens controllability; without Llin, FMD, CRA, and SRA degrade, and without Lmono, CRA and SRA drop.Source: §4.7 · p. 12
4.8 Qualitative Results
- Matched-view visualizations show clearer intensity separation while preserving motion identity, especially for motions with different amplitudes such as kick versus hop.The visual trend is consistent with higher SRA and lower LinErr and MonoViol in Tables 2 and 3.Source: §4.8 · p. 12
- The cross-content reference-based slider example illustrates practical animation editing, but is not claimed as a benchmark of fully general text- or arbitrary-reference transfer.The example goes beyond the strict paired evaluation setting.Source: §4.8 · p. 12
4.9 User Study
In a study of 11 participants across six matched motion cases, Motion Style Slider was rated higher for stylization strength and distinguishability than MCM-LDM, with comparable naturalness. The directional results were limited by the small sample size.
4.9 User Study
- The Likert-scale study used 11 participants and six motion cases, with anonymized video triplets at α values 0.5, 1.0, and 1.5.Participants rated stylization strength and naturalness for each video and distinguishability of intensity differences.Source: §4.9 · pp. 12–13, Table 5 · p. 14
- Stylization ratings were higher at low (3.92 vs. 3.56), mid (4.61 vs. 3.59), and high (5.18 vs. 3.46) intensity; distinguishability was 4.92 vs. 2.36.Naturalness was comparable, and monotonic-pass rates were 0.50 vs. 0.00; Table 5 summarizes six matched cases.Source: §4.9 · p. 13, Table 5 · p. 14
- Paired sign tests followed the same trend, but limited sample size left the reported differences with limited statistical significance.Reported p-values were 0.2188 for high-intensity style, 0.0625 for distinguishability, and 0.2500 for monotonic-pass improvement.Source: §4.9 · p. 14
5 Discussion
The discussion frames style intensity as a relative editing axis and distinguishes in-range interpolation from extrapolation measured against real stronger-style targets. It also identifies endpoint-pair quality, rare styles, sparse actors, and reference-based control as limitations.
5 Discussion
- SRA, LinErr, MonoViol, and ExtraErr directly assess requested direction and intensity ordering, whereas CRA and FMD support evaluation of content preservation and realism.Plausible individual samples do not by themselves show that a slider responds consistently to “less,” “more,” or extrapolated style.Source: §5 · p. 14
- Artists need a predictable relative control axis because style intensity is not an absolute scale shared across users.The quantitative evaluation remains endpoint-paired, while Figure 4 suggests reference-based control can also be useful across different content motions.Source: §5 · p. 14
- Against real stronger-style targets, methods with similar in-range recognition can differ in extrapolation error and monotonicity; our method's lower ExtraErr suggests a more meaningful out-of-range trajectory.The over-reaction benchmark evaluates against real motions at stronger style levels than the observed endpoints.Source: §5 · p. 15
- Control depends on endpoint-pair quality and style-embedding calibration; large-α artifacts may remain for rare styles or sparse actors, and evaluation focuses on reference-based endpoint control.The paper identifies unseen-style or text-guided control as future work.Source: Limitations · p. 15
6 Conclusion
Motion Style Slider combines endpoint-derived style directions, a scalar intensity control, and regularization for continuous control without intermediate-intensity supervision. Across three datasets, it improves control-oriented behavior while remaining competitive on realism and content metrics.
6 Conclusion
- The framework combines latent style-direction conditioning, a scalar control parameter, and regularization for linearity and monotonicity without intermediate-intensity supervision.It supports controllable interpolation and extrapolation from endpoint supervision.Source: §6 · p. 15
- Across PerMo, Bandai-Namco, and Xia, the method improves style fidelity and slider consistency while remaining competitive on realism and content metrics.The conclusion also reports an over-reaction benchmark using real stronger-style motions to measure out-of-range performance directly.Source: §6 · p. 15
- Future work includes stronger biomechanical constraints for high-α stability, improved cross-dataset style alignment, and larger user studies.Source: §6 · p. 15