Summary
Researchers from the Institute of Science Tokyo and Cygames have developed Motion Style Slider, a diffusion-based framework for continuously adjusting the intensity of style in generated human motion. Instead of requiring motion-capture examples at many expressive levels, the system learns from two endpoints, such as neutral walking and happy walking. A user-controlled value then moves the result from neutral through the target style and can extend toward stronger expression. The team reported favorable benchmark results for consistent intensity control, smooth transitions and extrapolation compared with existing methods. In a study of 11 university participants, viewers could distinguish intensity levels while rating naturalness as comparable. The work was presented at ECCV 2026 and published in Lecture Notes in Computer Science.
Hana’s perspective
I find this direction more useful than another system that simply generates a finished motion from a label. Animators work through degrees: less tired, more confident, hold the reaction back. Turning that language into a continuous control preserves iteration and gives a director something closer to a performance dial than a slot machine. The limited user study means the result is not yet proof of production readiness, and extrapolating beyond recorded styles could still produce unstable motion. But training from only neutral and styled endpoints could make nuanced controls practical without multiplying expensive capture sessions.
Are you a writer with a great story? We'd love to hear from you: APPLY HERE