Turn any AI-generated video into an animation for any rigged creature (with wings, tails, or even eight legs): the video model captures both front and side views in the same frame, a joint recognizer specific to that creature reads the joints from both perspectives, triangulation determines the 3D positions, and then the skeleton is applied and exported as a GLB file.
Stars 420 (+61 / 24h)
Reasoning
- The workflow has 8 steps: generate a textured mesh from a single image (Flux 3 for image generation, Tripo H3.1 for modeling) → manually place the skeleton on the mesh view and skin it in Blender → paint each limb a different color to prevent left-right swapping → use MiniMax H3 Max to generate a 5-second dual-camera video based on prompts → train a DINOv2 joint recognizer using random poses → read joints frame by frame → compute 3D positions from the two views → fit the skeleton and export a 24 fps GLB
- The author provides examples of a wolf, dragon, giant spider, and fire demon; the recognizer's average error on the held-out set is 2.9–9.4 pixels; the author states that each 5-second video costs about 0.26 dollars on fal, joint reading takes 10 seconds, fitting takes less than a minute, and no manual correction is needed
- Includes 6 Claude Code skills (overview, rigging, filming, recognizer, capture, LoRA), with instructions on how to distinguish good captures from bad ones; additionally, a MiniMax H3 dual-camera LoRA can be trained on fal, and the author claims that adding the LoRA reduces limb color swapping in the wolf example from 2 out of 5 clips to 0 clips (running only one seed per action)
- The method comes from @CanItCode, the idea for the H3 LoRA comes from @Graalitoo, and this repository is the author's organized, runnable version
Usage
- Requires an NVIDIA CUDA GPU (training the recognizer needs 24 GB of VRAM, while joint reading and fitting have much lower requirements), Python 3.11, Blender 5.x, and a fal Key; after git cloning, install torch 2.11 and dependencies according to the README, and write FAL_KEY into .env
- Try it for free first: the repository includes joint data for a dragon roaring clip, which can be fitted in about 20 seconds on the CPU using mocap.solve; then use tools/fetch_examples.sh dragon --readers --clips to download example meshes, recognizers, and clips, and run prep.sh, rig.sh, and capture.sh in sequence to get the complete GLB
- Claude Code: /plugin marketplace add Patch-gg/anymove, then /plugin install anymove@anymove
- To create your own creature: copy the closest example from assets/ (four-legged wolf, winged dragon, eight-legged spider, bipedal demon with wings and tail), replace model.glb, and run prep, rig, dataset (about 25 minutes), train (about 20 minutes), film, and capture in sequence
Risks
- Limitations listed by the author: eight-legged walking cannot be fitted (the two views disagree significantly on which leg is lifted), lying-down actions fail, the skeleton must be placed manually, and the video model may drift; ground contact only recognizes paw naming for wolves and dragons, and fur and cloth require additional bones
- The LoRA training clips are derived by reverse-engineering motion from AI videos and then rendering them, so the motion itself is AI-generated; training on fal for 1200 steps costs about 12 dollars and takes 2 hours, with prices recorded by the author in 10
- The code is Apache-2.0, but the MiniMax H3 open weights are under a community license, usage on fal is subject to fal's terms, and the included wolf LoRA is subject to the H3 community license rather than Apache-2.0; the repository was created on 27 10