Appearance
JCA generates a video and the 3D rig motion behind it
October catch-up. Paper published October 20, 2026. Published November 13, 2026.
NVIDIA researchers introduced Joints, Camera, Action! on October 20, a system that generates a video together with motion for a supplied 3D rig. A creator provides the skeleton and a text prompt, with an optional scene image or selected keyframes. The result includes joint motion that can be replayed on the asset. Paper and publication date.
An animation you can inspect from another view
JCA connects a motion model to a pretrained video generator so the two exchange information during generation. It also predicts where joints appear in the video, allowing a camera to be fitted to the 3D motion. That alignment helps place the animation in the depicted scene. NVIDIA's project explanation.
The demonstrations include animals and robots with different skeletons. They show several ways to work: produce motion from a prompt, guide it with image keyframes, recover motion from a video, or use existing motion to generate a video. In the keyframe example, playback pauses at the supplied images so you can compare the requested pose with the generated transition. Interactive examples.
Contact still needs an animator's attention
The paper identifies a limit that matters for game assets. JCA sees the surrounding scene through the video, without access to its 3D geometry. Nothing in that setup enforces foot contact or prevents a body from passing through an object. The camera alignment assumes a static camera, and each clip animates one rig. Objects that move during an interaction aren't recovered as separate animated 3D objects. Limitations in the paper.
Those boundaries make contact-heavy actions a useful place to inspect the results: feet on a floor, a hand on a handle, or a jump clearing an obstacle. Attractive footage alone can't establish whether the motion fits the collision geometry in your level.
As of November 13, the project page provides the paper and demonstration videos but no linked code or model download. This remains research to watch for an animation workflow, with cleanup and integration work still to establish.