NVIDIA Research has built a system that turns a text prompt into a video and the matching 3D animation data for a character, both at once. It is called Joints, Camera, Action!, or JCA, and it works with the skeleton (the “rig”) of any 3D character, human or not.
Ordinary video models output pixels. According to NVIDIA’s project page, a video can depict a monkey leaping over a crossbar, yet nobody gets that leap back as movement placed in a 3D scene. JCA does, and it also estimates a camera so the motion lines up with the picture.
Give it a starting image and the motion reacts to the surroundings. A few chosen keyframes steer the path. It runs in reverse too, turning video into rig motion. Training used nearly 200,000 pairs, built partly by placing animated 3D assets into generated scenes, plus human and robot footage.
NVIDIA says JCA follows text prompts more closely than pipelines that start from video, and it shows side-by-side clips against AnimateAnyMesh, Puppeteer, and ActionMesh. No scores appear.
For game and film studios, the test is whether the motion needs less hand cleanup than today’s tools, and the demos do not answer that.
Reported by NVIDIA Research on its project page for Joints, Camera, Action!, which carries no publication date.
