Motion synthesis via adaptive guidance experts

Oliver Hixon-Fisher; Jarick Francik; Dimitrios Makris

Motion synthesis via adaptive guidance experts

Oliver Hixon-Fisher
, Jarick Francik
, Dimitrios Makris

Kingston University

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

Existing methods for pose-centric motion synthesis often rely
on specialised architectural components to understand how motion changes
over time. In this paper, we propose a novel approach that incorporates
adaptive instance normalisation within a dedicated Adaptive Guidance
Expert, integrated into a Mixture-of-Experts (MoE) system. This design
enhances the model’s ability to capture temporal coherence in synthetic
motions. Our method achieves state-of-the-art performance in generating
realistic motion, as measured by Fréchet Inception Distance (FID),
while maintaining comparable diversity. We validate its effectiveness on
the CMU MoCap and HumanAct12 datasets, demonstrating its capability
to create plausible and high-quality motion sequences suitable for a
wide range of applications.

Original language	English
Title of host publication	Computer Vision, Imaging and Computer Graphics Theory and Applications
Publisher	Springer Nature
Publication status	Accepted/In press - 17 Sept 2025
Event	20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Porto, Portugal Duration: 26 Feb 2025 → 28 Feb 2025

Publication series

Name	Communications in Computer and Information Science
ISSN (Print)	1865-0929
ISSN (Electronic)	1865-0937

Conference

Conference	20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications
Period	26/02/25 → 28/02/25

Cite this

@inproceedings{1c72b73d80cf4f029f62f683b6edd982,

title = "Motion synthesis via adaptive guidance experts",

abstract = "Existing methods for pose-centric motion synthesis often rely on specialised architectural components to understand how motion changes over time. In this paper, we propose a novel approach that incorporates adaptive instance normalisation within a dedicated Adaptive Guidance Expert, integrated into a Mixture-of-Experts (MoE) system. This design enhances the model{\textquoteright}s ability to capture temporal coherence in synthetic motions. Our method achieves state-of-the-art performance in generating realistic motion, as measured by Fr{\'e}chet Inception Distance (FID), while maintaining comparable diversity. We validate its effectiveness on the CMU MoCap and HumanAct12 datasets, demonstrating its capability to create plausible and high-quality motion sequences suitable for a wide range of applications.",

author = "Oliver Hixon-Fisher and Jarick Francik and Dimitrios Makris",

year = "2025",

month = sep,

day = "17",

language = "English",

series = "Communications in Computer and Information Science",

publisher = "Springer Nature",

booktitle = "Computer Vision, Imaging and Computer Graphics Theory and Applications",

address = "Netherlands",

note = "20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications ; Conference date: 26-02-2025 Through 28-02-2025",

}

TY - GEN

T1 - Motion synthesis via adaptive guidance experts

AU - Hixon-Fisher, Oliver

AU - Francik, Jarick

AU - Makris, Dimitrios

PY - 2025/9/17

Y1 - 2025/9/17

N2 - Existing methods for pose-centric motion synthesis often rely on specialised architectural components to understand how motion changes over time. In this paper, we propose a novel approach that incorporates adaptive instance normalisation within a dedicated Adaptive Guidance Expert, integrated into a Mixture-of-Experts (MoE) system. This design enhances the model’s ability to capture temporal coherence in synthetic motions. Our method achieves state-of-the-art performance in generating realistic motion, as measured by Fréchet Inception Distance (FID), while maintaining comparable diversity. We validate its effectiveness on the CMU MoCap and HumanAct12 datasets, demonstrating its capability to create plausible and high-quality motion sequences suitable for a wide range of applications.

AB - Existing methods for pose-centric motion synthesis often rely on specialised architectural components to understand how motion changes over time. In this paper, we propose a novel approach that incorporates adaptive instance normalisation within a dedicated Adaptive Guidance Expert, integrated into a Mixture-of-Experts (MoE) system. This design enhances the model’s ability to capture temporal coherence in synthetic motions. Our method achieves state-of-the-art performance in generating realistic motion, as measured by Fréchet Inception Distance (FID), while maintaining comparable diversity. We validate its effectiveness on the CMU MoCap and HumanAct12 datasets, demonstrating its capability to create plausible and high-quality motion sequences suitable for a wide range of applications.

UR - https://link.springer.com/series/7899

M3 - Conference contribution

T3 - Communications in Computer and Information Science

BT - Computer Vision, Imaging and Computer Graphics Theory and Applications

PB - Springer Nature

T2 - 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications

Y2 - 26 February 2025 through 28 February 2025

ER -

Motion synthesis via adaptive guidance experts

Abstract

Publication series

Conference

Other files and links

Fingerprint

Cite this