patch-gg/anymove

This is a working pipeline rather than a polished product: it can produce creature mocap from an AI clip, but expect to hand-tune rigs and accept several motion types that simply fail.

The problem

Getting believable motion onto a non-human character usually means either buying or shooting reference, then keyframing or retargeting skeleton by hand for a creature with wings, a tail or eight legs. That loop is slow, and most mocap libraries have nothing that resembles a dragon. anymove changes the input side: you generate a clip in which a video model shows the creature from front and side in one frame, then the tool locates the joints in both views, combines them into 3D, fits your rig frame by frame and writes a GLB. The output is animation you can take into Blender, not a finished shot.

What's different

The mechanism is the interesting part. A joint reader built on DINOv2 is trained per creature, on a dataset of roughly 2,200 random poses rendered in both views, so it learns where that creature's particular joints appear. Because a joint sits at the same height in the front and side view, the pair of detections pins it in 3D. The rig is then fitted to those points, and the README reports how well the two views agree for each clip, which is a useful signal for whether a take is usable. Limb colours painted per limb prevent left/right swaps during fitting.

The cost of this design is that the reader is creature-specific: a new model requires placing joints by hand, generating a dataset, training for roughly 20 minutes on a 12 GB card, and then checking it. The two-camera framing is also not a standard video-generation output; it depends on a video model producing a consistent creature from two angles in one frame, and the README notes that plain MiniMax H3 swapped limb colours on 2 of 5 wolf clips (using a LoRA got that to 0 of 5). Cost and speed figures in the README ($0.26 per clip, under a minute to read and fit) are the author's numbers from October 2026, not independent benchmarks.

Who should try it

Good fit: a small game or animation team that already models creatures in Blender and is comfortable with Python, GPU setup and hand-placing skeleton joints on grid views. The README's example results are promising for large, single gestures: rearing, roaring, slamming, take-off and pouncing, with agreement scores like 0.99 for the spider rearing up and 1.00 for the dragon taking off.

Poor fit: anyone who needs locomotion or ground contact. The README is direct that eight-legged walking disagrees across views (0.36–0.49) and that lying down failed because the training poses never included it. If the target creature is a quadruped that mostly walks, or a character that needs to lie down, this will not carry the shot yet. The README also ships a Claude Code plugin with dedicated skills for rigging, filming, training and capture, which changes the entry point for people who work inside that assistant, but it is one path, not a requirement.

Requirements and limitations

Setup is the real gate. You need an NVIDIA GPU with CUDA, Python 3.11 and Blender 5.x on your PATH, plus a fal key to generate new clips; reading and fitting need far less than the 24 GB that training does. Blender is not optional for the full pipeline unless you supply a prepared mesh and skeleton. The example release carries a dragon mesh, a 355 MB trained reader and a clip, so you can run the fit on CPU in about 20 seconds and see the two-view agreement before committing to your own creature.

Ongoing limits: skeletons are placed by hand on grid views; the reader only knows poses it was trained on; video models drift and can swap limb colours; and licensing is mixed. The repo is Apache-2.0, but the released wolf LoRA inherits MiniMax H3's community licence, and hosted fal generation runs under fal's terms, so check both before shipping.

HotRepo's take

The only dated signal is a NEW-REPOSITORY board (新项目榜) snapshot on 2026-10-27 showing 141 total stars at #25 for Python — a snapshot, not a daily growth figure. The repo was created that same day, so there is no trend history to judge persistence. Decide from the README's limits: if your creature stands, roars or slams and your team will hand-place joints, the pipeline is worth a test on the provided dragon example.

Project source and documentation