AI Motion Capture for Dragons and Spiders: anymove Turns a 26-Cent Video Into a Rigged Animation
Nobody has ever put a dragon in a motion capture suit. That is the running joke at the centre of creature animation: the tools built for capturing movement assume two arms, two legs and a human willing to wear dots. Anything with wings, a tail or eight legs gets keyframed by hand, one pose at a time.
A small studio called Patch just released a method that sidesteps the suit entirely. anymove takes a five-second AI video of your creature, reads where every joint sits, and fits your rig to it frame by frame. It exports a standard GLB animation file that any game engine or Blender can open. The authors put the video at about 26 cents per move.
The clever part is not the AI video. It is how the video is framed. And that framing tells you a lot about where AI clips are heading for people who make things: away from being the finished shot, toward being raw material you measure.
What anymove actually does
Think of it as motion capture where the actor is a video model.
You start with your own creature: a textured 3D model with a skeleton inside it. anymove renders that creature standing in a plain grey studio, seen by two cameras at once. The front view sits on the left half of the frame. The side view sits on the right. That still image goes to a video model along with a one-line description of the move, something like "the dragon whips its tail from side to side, then stands still."
The video model animates both views at once. Then a small vision model, trained on that one creature, finds every joint in both halves of every frame. A solver bends your rig to match and writes the animation out.
The repository credits the two-camera method to a YouTube creator, CanItCode, and the idea of training a small add-on on top of the video model to Graalitoo on X. Patch's contribution is a working version of the whole pipeline with step-by-step instructions, released under the Apache 2.0 licence on October 27.
Why two views in one frame is the whole trick
A single video gives you a flat picture. You can see where a dragon's nose is left to right and up and down, but not how far toward or away from the camera it is. That missing depth is why a normal clip cannot become 3D motion.
anymove solves this with geometry a draughtsman would recognise. Both cameras share the same height scale and the same floor line. So a joint at a given height sits on the same row in both halves of the frame. The front view tells you how far left or right the nose is. The side view tells you how far forward or back. Height comes from both.
The README describes it as standing the two views up like walls at a right angle. Each joint is where its two sight lines cross.
There is a second detail worth stealing even if you never use the tool. Before filming, every limb gets painted its own flat colour. The video model then animates a creature with a red front-left leg and a blue front-right leg, and the joint reader can never confuse them. Left and right swapping is the classic failure of reading motion from video. Colour coding the limbs is a low-tech fix for it.
What it costs, honestly
The 26 cents is the price of each new move. It is not the price of starting.
Here is the full bill from the README, at fal's October 2026 prices:
- A 3D model of the creature: about $0.40 if you generate one with Tripo, or free if you already have one.
- A skeleton: you place the joints by hand on grid views. This is real work and it is yours.
- Training the joint reader: about 25 minutes of rendering 2,200 random poses, then 17 to 20 minutes of training on one graphics card. The quick start asks for an NVIDIA card with 24 GB of memory for this step.
- Each move: about $0.26 for a five-second clip from MiniMax H3 Max on fal, then about 10 seconds to read it and under a minute to fit the rig.
So the honest framing is: one afternoon of setup per creature, then animations at the price of a coffee bean each. For a game with one hero creature that needs twenty moves, that trade is excellent. For twenty creatures with one move each, it is not.
Put this into practice
You need an NVIDIA graphics card, Python 3.11, Blender 5.x and a fal account for making new clips. The repository also works as a Claude Code plugin, with a separate skill for each stage, if you prefer a coding assistant to walk you through it.
Try it with no graphics card first. The repository ships the joints its reader already found in a dragon's roar. One command fits the rig to them on an ordinary computer in about 20 seconds and prints how well the two views agree. Do this before you spend anything. It shows you what the output looks like and what "good agreement" means.
Pick the nearest example creature. The repo includes four: a wolf (four legs), a dragon (four legs and wings), a giant spider (eight legs) and a fire demon (two legs, two arms, wings and a tail). Copy the closest one and put your model in its folder. Starting from a working setup saves you from inventing the configuration from scratch.
Prepare and rig. tools/prep.sh scales your model to real size and turns it to face the right way. Place the joints on the grid views, preview with tools/rig.sh <name> preview, then run tools/rig.sh <name> to build the skeleton, attach the mesh and paint the limbs.
Teach the reader. tools/dataset.sh renders the random poses and tools/train.sh trains the reader. On poses it never saw in training, the authors report the reader's average miss at roughly 3 to 9 pixels across their four creatures.
Film and capture. Write the move with tools/make_prompt.py, generate it with tools/film.sh, and turn it into an animation with tools/capture.sh. You get animation.glb plus a side-by-side video that lets you compare the result against the clip.
Choose moves that suit the method. The authors' own advice: big single gestures work best. Rearing, roaring, slamming, taking off, pouncing. Start there.
Where it breaks
Complex footwork fails. Eight-legged walking does not fit: the two views disagree about which legs are lifted. The spider works for rearing up, biting and pouncing, not for scuttling.
Lying down fails. The reader only learns poses inside the skeleton's normal limits, and lying on a side is outside them. The wolf's death and lie-down clips did not work.
Video models still drift. Plain MiniMax H3 swapped the limb colours on 2 of 5 wolf clips. A small add-on the authors trained on top of the video model, using the two-camera layout, brought that to 0 of 5. They are careful to say that is one try per move, a first read and not a benchmark. Expect to film some moves twice.
Fur, cloth and ground contact need extra work. Floppy parts need spring bones or tracked tips. The solver's sense of feet touching the ground only knows the wolf's and dragon's paw names, so other creatures fit without it and may slide.
The licences stack up. The anymove code is Apache 2.0. The clips come from MiniMax H3, which has its own community licence, and running it on fal adds fal's terms. The repository says to check both before shipping anything you make. The authors' wolf add-on is covered by MiniMax H3's licence, not Apache. Read those terms before a client sees your animation.
The motion is AI-made. It looks plausible, which is not the same as being physically accurate. For a fantasy creature that is usually fine. For a real horse in a documentary, keyframe it or license real capture.
The bigger shift: clips as measurements
Step back from dragons for a moment.
Most of the talk around AI video is about the shot: how real it looks, how long it runs, whether the hands are right. anymove does not care about any of that. Nobody watches its clips. The video exists to be measured, and the only quality that matters is whether the two views agree with each other from frame to frame.
That is a different way to judge a video model, and it is spreading. This same week, another open project turned short AI clips into playable browser games by treating the video as the design document. When the clip becomes a measuring instrument, consistency beats beauty, and a 26-cent render is worth more as data than as footage.
If you animate creatures, start with the free roar fit. Watch the side-by-side. If the agreement score is close to 1 (the README's scale, where 1 means both views show the same footfalls) and the motion reads, you have just seen what an afternoon of setup could buy you for every move your creature will ever need.