Why motion inbetweening matters

If you've ever watched a behind-the-scenes reel of an animated film, you've seen the division of labour: senior animators draw the key poses, and junior animators fill in the frames between them. Those in-between frames are called "inbetweens," and they're what make motion feel continuous.

In 3D animation, the process is different in tooling but identical in principle. An animator sets keyframes on a timeline — the character's pose at specific moments. The software interpolates between them. But interpolation is mechanical. It doesn't understand weight, momentum, or style. So the animator spends hours (or days) correcting the result, adding breakdowns, adjusting curves, and layering in the subtlety that makes motion feel real.

This is where most of an animator's time actually goes. Not on the creative decisions — those happen at the keyframes. But on the execution between them.

We think AI can change this equation without changing the animator's role.

The key insight is that motion inbetweening is a structured generation problem. You have clear constraints (the key poses), clear physics (gravity, momentum, joint limits), and a rich space of valid solutions in between. This is exactly the kind of problem that modern generative models — diffusion models, in particular — are good at.

But generating plausible motion isn't enough. The motion has to respect the animator's intent. It has to match the style of the production. It has to be adjustable. This is where most academic approaches stop and where we're focused: making the generation directable.

When we show our tool to working animators, the reaction is always the same. They're not impressed by the AI generating motion. They're impressed that they can change it — and that it still works.

That's the bar we're building to. Not "look what AI can do" but "look how fast you can work."