patch-gg, the team behind trompe, released anymove under the Apache-2.0 license on October 27, according to the project's GitHub repository. The tool captures motion for any rigged creature from a single AI-generated video, and it had about 380 GitHub stars by October 29.
Existing tools serve human motion capture well, but creatures with unusual body plans, such as wings, tails or eight legs, usually need hand animation.
The process starts with a video model. MiniMax H3 Max, run through fal, produces a five-second clip for about $0.26 that shows the creature from the front and the side in the same frame.
anymove then trains a DINOv2-based joint reader for that one creature, using 2,200 random poses. Training takes about 20 minutes on one GPU and cuts the average joint error from 40 pixels to 6. The reader finds every joint in both views, and the two views fix each joint's position in 3D.
The rig is fitted frame by frame and exported as a 24 fps GLB with no hand edits, in under a minute per clip. To create the creature itself, the example pipeline uses a FLUX 3 image and a Tripo H3.1 mesh for about $0.40, plus a 42-bone Blender rig.
The README credits the method to @canitcode, whose Opus 5.5 and GPT Astra Unreal Engine 5 boss-fight demo is circulating on X, and credits @graalitoo with the MiniMax LoRA idea.
The hardware bar is high. anymove needs an NVIDIA CUDA GPU with 24 GB to train the joint reader, plus Python 3.11, Blender 5.x and a fal API key. A CPU demo fit is included for people without that hardware.
Why it matters: animation has been one of the hardest parts of AI-assisted game production to automate, especially for non-human characters. Getting usable creature motion from a cheap generated video could save small teams a lot of time.
Caveat: the costs, timings and error figures come from the project's README, and AIGamer hasn't tested the tool.
What to watch: how patch-gg builds out anymove and trompe, which handles textures, as a set of tools for finishing blockout-stage prototypes.
