anymove: Open-Source AI Video and DINOv2 Motion Capture Tool for Non-Humanoid 3D Creatures

Patch has released 'anymove', an Apache 2.0-licensed motion capture pipeline that turns one AI-generated video into animation for a rigged creature. The repository demonstrates dragons, giant spiders, wolves, and other rigged creatures, including creatures with wings, tails, and eight legs.

Image source: https://github.com/patch-gg/anymove

The project pairs split-frame dual-camera video generation with a per-creature fine-tuned vision model, turning prompt-driven creature clips into 3D animation without manual keyframing.

Dual-Camera AI Video Generation and DINOv2 Joint Tracking Pipeline

The core mechanism of anymove combines two-view framing with 3D coordinate reconstruction.

The video model places a front view on the left and a side view on the right in one frame and animates both views at once from a prompt. Meta's DINOv2 vision model is then fine-tuned on 2,200 synthetic random poses rendered from the target creature's textured mesh over 20 epochs to create a specialized joint reader.

• Tracking Precision: The repository's training example reports that average miss falls from 40 pixels to 6 pixels; held-out front and side reader errors vary by creature.
• 3D Coordinate Reconstruction: The pipeline derives the X coordinate (lateral) from the front camera and the Y coordinate (depth) from the side camera, using shared Z coordinates (elevation) to reconstruct joint positions in 3D space.
• Automated Rig Fitting and Export: Bones from a Blender rig bend to match the tracked 3D points frame by frame, ignore stray coordinates, and export the animation as a 24 fps GLB for use in engines.

Hardware Requirements, Setup, and Claude Code Agent Skills

Training a custom joint reader requires local CUDA acceleration, while generating new clips requires an external API key.

Training a custom joint reader requires an NVIDIA GPU with 24 GB VRAM; reading and fitting need much less memory, and the included dragon fitting example runs on a CPU. The local environment requires Python 3.11 and Blender 5.x configured on the system PATH.

git clone https://github.com/Patch-gg/anymove && cd anymove
python3.11 -m venv .venv
.venv/bin/pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cu128
.venv/bin/pip install -r requirements.txt
cp .env.example .env

Video generation relies on the MiniMax H3 Max model via the fal platform, requiring a FAL_KEY in the local .env file.

anymove also packages dedicated agent skills for Claude Code. Through its plugin structure (anymove, anymove-rig, anymove-film, anymove-reader, anymove-capture, and anymove-lora), the packaged agent skills cover phases from mesh scaling and rig painting to prompt generation, reader training, capture checks, and showcase rendering.

Performance Benchmarks and Operational Costs

The repository reports benchmarks across several creature models: a wolf (36 joints) achieved a 0.94 trot agreement and 0.94 cm mean 3D joint error, while a dragon (50 joints) scored 1.00 on take-off and 0.93 on roar animations.

Reported costs include roughly $0.40 for a Tripo H3.1 mesh and around $0.26 for a 5-second two-camera MiniMax H3 Max clip on fal ($0.30 on H3); fitting takes under a minute per clip.

Reported agreement scores vary by creature and motion, so each generated clip should be checked before capture.

Sources

• GitHub: patch-gg/anymove
• Rehan Sheikh on X (@rehan_shei): 2026-10-27 anymove Release Announcement