Prompt2rig turns a text prompt into 3D motion on a Mixamo rig

Animators have had two ways to get a short animation. One is to hire a mocap studio. The other is to hand-animate every joint. A tool called prompt2rig does neither. The user types a short description of a motion. A video model acts it out. The tool then extracts the 3D body motion. It plays that motion on a rigged Mixamo character in the browser.

The project is open source. The universal-modder crew published it on 19 October 2026. The licence is Apache-2.0. It has 407 stars at the time of writing. GitHub shows the repository was created six days ago.

How the three steps work

The prompt and a performer image go to a video model on fal. The author lists two choices. FLUX 3 Max gives the most realistic motion. It takes about 8 to 9 seconds for a 5 second clip at 720p. H3 Max Turbo is the fastest. It takes about 2 to 3 seconds at 480p.

In the browser, MediaPipe tracks the pose on each frame of the clip. It plays the result live on the Mixamo character. The author calls this a preview, so the user is not waiting on the slower pass.

The slower pass runs on a local Windows machine with an NVIDIA GPU. GVHMR reads the whole clip. It outputs clean 3D joint rotations in about 3.5 to 4 seconds. WiLoR runs at the same time. It adds wrists and the 15 finger joints per hand in another 0.3 seconds. The user can export the final take as a .glb file.

What works and what does not

The author says the tool is best at motions that face the camera. The limbs should stay clear of the body. Dances, kicks, kata, sport mimes, celebrations and clear hand shapes all work. Flips, spins, floor work and crossed arms confuse a single-camera pose estimator. The presets steer away from them.

The licence picture is split. Prompt2rig's own code is Apache-2.0. The web app, the live preview and video generation are commercially usable. The slower pass uses GVHMR, SMPL-X, WiLoR and MANO. The author lists each one as research or non-commercial only. A commercial product using that path would have to replace or license each piece.

Takes export to the same Mixamo rig the user already knows. The rig works in Blender and Unity. So an animator can take the rough result and clean it up in a tool they already use.

For a team building creative tools, the small size of this repository is itself the point. One person-weekend of work ties together an existing video model, two research pose estimators and a browser retargeting layer. The output is close enough to be useful on day one.

Source

blendi-remade/prompt2rig on GitHub.