MorphoStyle: Motion Style Transfer with Morphology Control

Xin Feng; Eleonora D'Arnese; Mohan Sridharan

cs.CVSep 12, 2026

MorphoStyle: Motion Style Transfer with Morphology Control

Authors: Xin Feng, Eleonora D'Arnese, Mohan Sridharan

Organizations: Institute of Perception, Action, and Behavior School of Informatics University of Edinburgh Edinburgh, UK

Abstract

Human motion may be viewed as a combination of action content, style, and body morphology. Existing motion style transfer methods transfer a reference style onto a content motion while assuming a canonical body, whereas shape-aware motion generators adapt motion to a target shape without explicit style control. This separation of motion style and shape (morphology) makes it difficult to generate stylized motions for non-canonical bodies; naively combining a style transfer module with a shape-aware generator often leaks action content from the style reference and disrupts shape-consistent kinematics. We present MorphoStyle, a framework for shape-aware motion style transfer that is built on a shape-conditioned Finite-Scalar-Quantization Variational Auto-Encoder (FSQ-VAE). The key contribution is to pose the desired style transfer as modular latent disentanglement comprising: (i) a contrastive style encoder that extracts content-decoupled style embeddings; (ii) a text-guided style-routing mechanism that locates style-relevant joints in a text-motion feature space; and (iii) a manifold preserving style modulator that injects discriminative style embeddings in content features as a temporally-gated low-rank offset. Extensive experiments on benchmark datasets demonstrate that MorphoStyle outperforms competing baselines in terms of both shape control and motion style transfer, while simultaneously providing quantitative shape control. For more details, please see project website: https://github.com/funkdub/MorphoStyle.

Figures & tables

Figure 1: \textcolor [rgb]0,0,0 Framework Overview. Given a content motion $x_{c}$ , a style reference $x_{s}$ , and a textual body-shape description $t_{s}$ , MorphoStyle synthesizes an output motion $\hat{x}$ that preserves the action of $x_{c}$ , transfers the style of $x_{s}$ , and matches the target body shape. A frozen shape-aware FSQ-VAE tokenizes $x_{c}$ into content codes $c$ and decodes $\hat{x}$ via a shape-conditioned decoder $\mathcal{D}$ . A trainable style branch produces a global style embedding $s_{g}$ and a joint-localized embedding $s_{j}$ , fuses them into style code $s$ , and injects $s$ in $c$ through a Manifold-Preserving Style Modulator (MPSM) as a temporally-gated low-rank residual. A frozen Shape-to-Feature module (S2FM) maps $t_{s}$ to shape embedding $\phi(\beta)$ conditioning $\mathcal{D}$ .

Figure 2: \textcolor [rgb]0,0,0 Structure of Manifold-Preserving Style Modulator (MPSM).

Figure 3: \textcolor [rgb]0,0,0 Qualitative comparison of shape-aware motion style transfer against SMooDi [ 43 ] and SMooDi+ShapeMyMove. Yellow body is the target shape used for reference.

Method	SRA $\uparrow$	FID $\downarrow$	Foot Skate $\downarrow$	Shape Control
MLD + Aberman et al	61.1%	3.89	0.338	$\times$
MLD + Motion Puzzle	67.2%	6.87	0.197	$\times$
SMooDi	65.1%	2.95	0.095	$\times$
StyleMotif	68.8%	1.38	0.094	$\times$
ShapeMyMove	1.57%	1.36	0.228	$\checkmark$
SMooDi + ShapeMyMove	69.9%	7.87	0.245	$\checkmark$

Table 1: Comparing MorphoStyle with baselines for motion style transfer on HumanML3D content with 100Style references. Bold indicates best result and underline the second-best.

Method	BoneLen-MAE $\downarrow$	BoneLen-T-Std $\downarrow$	$\beta$ -swap $\uparrow$	$\beta$ -corr $\uparrow$
ShapeMyMove	4.59	3.53	5.05	0.677
SMooDi + ShapeMyMove	2.49	3.31	3.87	0.314
MorphoStyle (ours)	1.91	1.26	5.31	0.332

Table 2: Comparison with competing methods for shape control during motion style transfer using four evaluation measures. Bold indicates the best result and underline the second-best.

Method	Trainable params	Inference time $\downarrow$	GFLOPS $\downarrow$
SMooDi	13.90 M	7.19s	79.59
SMooDi+ShapeMyMove	31.11 M	2.11s	112.84
MorphoStyle (ours)	4.25 M	0.02s	1.98

Table 3: Comparison of model parameters, inference speed, and computational cost.

Figure 4: t-SNE visualization of global style embeddings $s_{g}$ on six representative style classes, with (right) versus without (left) the contrastive supervision $\mathcal{L}_{\mathrm{cl}}$ .

Variant	Contrastive	Joint	FID $\downarrow$	SRA $\uparrow$	Foot $\downarrow$
Base	$\times$	$\times$	3.57	74.7%	0.143
JTC	$\times$	$\checkmark$	3.54	78.5%	0.139
Con	$\checkmark$	$\times$	3.05	81.2%	0.110
MorphoStyle	$\checkmark$	$\checkmark$	2.84	81.9%	0.081

Table 4: Ablation on content-style disentanglement components.

Figure 5: Qualitative ablation of two training objectives, $\mathcal{L}_{\mathrm{dis}}$ and $\mathcal{L}_{\mathrm{shape}}$ , on the same content (walk forward), style reference ( kick ), and target body shape.

	SRA $\uparrow$	FID $\downarrow$	BoneLen-MAE $\downarrow$	$\beta$ -swap $\uparrow$
w/o MPSM	62.9%	5.16	1.97	2.06
w/o gate $g$	72.1%	4.85	1.92	4.12
w/o $\mathcal{L}_{shape}$	80.4%	2.99	2.20	3.41
MorphoStyle	81.9%	2.84	1.91	5.31

Table 5: Ablation results on key technical components of our MorphoStyle.

Finite Scalar Quantization

See all 11 papers

Sep 7, 2026

Flexible Motion Generation from Language and Style References

We introduce FlexMoGen, a novel framework for flexible human motion synthesis conditioned on both natural language descriptions and motion style references. Text prompts are effective at defining semantic content, but they are often limited in capturing fine-grained style details such as timing, limb articulation, and expressive dynamics. A style example clip supplements the text by conveying these nuanced motion characteristics directly, enabling the model to preserve high-level intent while reproducing the desired stylistic traits. Given a text prompt and a style example clip, FlexMoGen generates high-quality motions that preserve semantic content while faithfully reflecting the target style, offering users greater control over the animation generation process. Unlike prior methods that rely on discrete style labels and do not generalize to long or multi-style generation, FlexMoGen learns a variational style encoder without style supervision and supports long, time-varying, multi-style synthesis. Our framework jointly pre-trains the style encoder and a text-to-motion latent diffusion model within a unified architecture, modulating motion style through a lightweight adaptation module. It integrates an efficient relative positional encoding scheme and is trained on both stylized and non-stylized datasets, enabling strong generalization to unseen text-style combinations. Experiments show that FlexMoGen achieves the best balance between content fidelity and style reflection.

Kai Weixian Lan, Bodie Criswell, Briana Fedkiw +3

University of California, Davis, USA · University of Utah, USA · Epic Games, USA +1

Text-to-Motion Generation Motion Diffusion Models Human Motion Generation Style-Conditioned Motion Generation Diffusion Models for Motion Generation Diffusion Models

Ask PDF

Sep 3, 2026

STyMo: Fast and Controllable Few-Shot Motion Style Transfer

Supporting a wide variety of motion styles is critical for creating diverse virtual characters, but current methods either require large stylized datasets or pre-trained models that cannot generalize beyond their training distribution. We present STyMo, a few-shot approach that learns motion style from only seconds of paired data and trains in one to two minutes. Our key insight is to decompose style into two components: a static component capturing time-invariant posture, and a temporal component capturing frame-wise dynamics. This decomposition yields an interpretable system where posture intensity, temporal exaggeration, and per-body-region style can be adjusted at runtime. Furthermore, the reduction in required training data and computation time structurally permits an iterative authoring workflow. To ensure robustness on arbitrary inputs, we further introduce a stylizability gate that automatically prevents artifacts on out-of-distribution motions. We demonstrate results across diverse motion styles, from subtle emotional variations to exaggerated character archetypes, and release our processed paired dataset to facilitate future research.

Jose Luis Ponton, Alexander Winkler, Ladislav Kavan +2

Reality Labs, Meta, Zurich, Switzerland · Universitat Politècnica de Catalunya, Barcelona, Spain · Reality Labs, Meta, Sausalito, USA +2

Few-Shot Learning Motion Representation Learning

Ask PDF

Sep 25, 2026

Motion Style Slider: Endpoint-Supervised Continuous Style Control for Human Motion Diffusion

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.

Chen-Chieh Liao, Yichen Peng, Yiyi Cai +5

Institute of Science Tokyo · Cygames, Inc. · The University of Tokyo

Human Motion Generation Diffusion Models for Motion Generation Motion Representation Learning

Ask PDF

Abstract

Finite Scalar Quantization

Explore similar work

Flexible Motion Generation from Language and Style References

STyMo: Fast and Controllable Few-Shot Motion Style Transfer

Motion Style Slider: Endpoint-Supervised Continuous Style Control for Human Motion Diffusion