Motion Style Slider

Continuous style-intensity control for human motion diffusion.

ECCV 2026

Motion Style Slider

Endpoint-Supervised Continuous Style Control for Human Motion Diffusion

Chen-Chieh Liao1,3,*, Yichen Peng1, Yiyi Cai2, Yûi Ono3, Hiroki Hanaoka3, Erwin Wu1, Hideki Koike1, and Shuichi Kurabayashi3

1Institute of Science Tokyo    2The University of Tokyo    3Cygames, Inc.

*The work was partially done during an internship at Cygames, Inc.

Paper
Motion Style Slider changes a human motion smoothly from neutral to stylized and beyond the observed endpoint.

A slider for motion style

Motion Style Slider is a motion-to-motion diffusion framework for fine-grained control of style intensity. Given a content motion and a stylized endpoint, the model generates a continuous family of motions controlled by a scalar value, from neutral behavior through the target style and into stronger out-of-range reactions. The control is relative and monotonic, matching the way artists and directors naturally ask for a motion to be “a little more” or “a little less” expressive.

Continuous One scalar controls smooth style intensity instead of selecting only discrete labels.
Endpoint-supervised Training does not require ground-truth motions at every intermediate intensity.
Extrapolatable The model supports controlled over-reaction beyond the observed style endpoint.

How it works

We represent the change from content motion to style motion as a direction in a learned motion-style embedding space. Moving along that direction with intensity \(\alpha\) creates the style condition for a pretrained motion diffusion denoiser. Diffusion training is combined with latent intensity regularization to encourage smooth, ordered changes while preserving the original action.

Motion Style Slider method pipeline.
The model builds a style direction from endpoint motions, samples an intensity, and conditions the diffusion denoiser with separate content and style signals.

Qualitative results

Across several motion-style datasets, the method aims to preserve the source action while changing expression predictably as the slider moves. Evaluation covers realism, content preservation, monotonicity, interpolation, and extrapolation to newly captured over-reaction targets.

Qualitative comparison of Motion Style Slider with motion style transfer baselines.
Qualitative comparisons with representative motion-style transfer baselines at multiple style intensities.
Cross-content qualitative results for Motion Style Slider.
Cross-content stylization examples showing the same style-control mechanism across different source motions.

Citation

@inproceedings{liao2026motionstyleslider,
  title     = {Motion Style Slider: Endpoint-Supervised Continuous
               Style Control for Human Motion Diffusion},
  author    = {Liao, Chen-Chieh and Peng, Yichen and Cai, Yiyi and
               Ono, Yûi and Hanaoka, Hiroki and Wu, Erwin and
               Koike, Hideki and Kurabayashi, Shuichi},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}