IBMM++ gives animators finer control over AI-made motion
Animators get finer control over AI motion
IBMM++ is an extended version of the Implicit Bézier Motion Model, which lets animators set motion keyframes at any timing and now also extends and edits long sequences.
Diffusion models can generate diverse movements from high-level signals such as text. Motion diffusion that uses Bézier curves has been shown to allow precise joint-level conditioning. Earlier Bézier-based versions were limited to joints at a fixed temporal stride.
The original work, by Vögeli et al., trained the model on all possible configurations of Bézier control points. That enabled control at arbitrary timings. Animators require more temporal flexibility when keyframing or manipulating tangents to achieve principles such as easing in and out. The original also introduced a quantitative measure of ease-in and ease-out, which leads to a novel condition over the motion generation process.
IBMM++ builds on this. According to the description, it improves constraint accuracy and reduces foot sliding. It also enables motion extension, orientation constraints, longer sequences, and stable iterative editing. The description attributes this to architectural updates, a revised tangent formulation, a randomized constraint sampling strategy, and dedicated inference-time methods. It was validated on a larger internal dataset with more varied motions.
A video from the DisneyResearchHub channel walks through the work.
Would you rather keyframe by hand or let a model fill in the motion?
Video: DisneyResearchHub on YouTube