Science Tokyo and Cygames Introduce AI Slider for Fine-Tuned Character Motion

Researchers at the Institute of Science Tokyo and Japanese video game developer Cygames have developed an artificial intelligence framework designed to give animators continuous control over character movement nuances without requiring extensive motion-capture recordings.

The system, named the Motion Style Slider, was developed by specially appointed assistant professor Chen-Chieh Liao and Professor Hideki Koike alongside Cygames engineers. The team presented the work on Sept. 10, 2026, at the European Conference on Computer Vision (ECCV 2026) and published their findings in Lecture Notes in Computer Science, as reported by TechXplore .

Modern digital entertainment production relies heavily on fine adjustments to character movement, such as making an action slightly more restrained or expressive. While commercial generative AI systems can transform neutral character motion into stylized behaviors like joy, anger, or exhaustion, existing tools typically generate fixed-intensity outputs. Creating datasets covering every intermediate emotional intensity has remained cost-prohibitive because of the extensive motion-capture studio time required.

The Motion Style Slider addresses this limitation by requiring only two reference endpoints: an unstylized neutral action and a single stylized counterpart, such as neutral walking and happy walking. The framework computes a directional vector between both states within a learned motion-style embedding space without depending on static style labels. Animators then adjust a scalar value, designated as α—such as 0, 0.5, 1.0, or 1.5—which combines with content data and feeds into a pretrained diffusion-based motion generation model.

This scalar mechanism allows animators to modulate movement along a smooth gradient or extrapolate past the initial demonstration into exaggerated expressions. In comparative testing, the researchers evaluated the architecture against existing approaches, including DeepMotionEditing and the Multi-condition Motion Latent Diffusion Model. Liao reported favorable performance in consistent style modulation, smooth transitions, and expressive extrapolation.

To test output fidelity, the researchers ran a Likert-scale user evaluation with 11 university participants. Reviewers rated the synthesized style variations as clearly distinguishable while noting that physical naturalness was preserved across the generated motions.