anymove: creature mocap from one two-view AI video

The trick. Human mocap from video is a solved-ish problem; wings, tails and eight legs aren't. anymove gets around it by asking a video model to film the creature from two cameras in one frame: front on the left, side on the right. A joint reader trained on that one creature finds every joint in both halves. Because a joint sits at the same height in both views, the pair pins it in 3D, and a solver bends the rig onto those points frame by frame and exports a GLB at 24 fps. The repo credits the two-camera method to @CanItCode and the MiniMax H3 LoRA step to @Graalitoo; Patch's version packages both with agent skills.

The pipeline, per the README. A FLUX 3 picture becomes a textured mesh in Tripo H3.1 (about $0.40). You place a skeleton on grid views and skin it in Blender, and every limb gets its own colour so left and right can't swap. MiniMax H3 Max on fal makes a 5-second two-camera clip for about $0.26. A DINOv2 reader learns that creature from 2,200 random rendered poses (about 25 minutes of renders and 20 minutes of training on one RTX 4090, 12 GB). Reading a clip takes about 10 seconds and fitting it under a minute. There are six Claude Code skills (rig, film, reader, capture, LoRA and an overview) that walk an agent through each step and its quality checks.

Results and limits, in the creator's words. The README reports reader errors of roughly 3 to 9 pixels on a 720-row frame across a wolf, dragon, spider and fire demon, and shows where it breaks. Eight-legged walking doesn't fit because the two views disagree on which legs are up. Lying-down poses fail. Skeletons are placed by hand. Plain H3 swapped limb colours on 2 of 5 wolf clips, versus 0 of 5 with the optional LoRA (about $12 on fal), which the authors call a first read, not a benchmark. Big single gestures (rearing, roaring, slamming, taking off) work best. Rehan's follow-up made 10 T-rex animations for about $4 and says they're 'not 100% there yet'.

What you need. An NVIDIA GPU with CUDA (24 GB to train a reader, much less to read and fit), Python 3.11, Blender 5.x and a fal key for new clips. You can try the fit step on a CPU with the bundled dragon example. The code is Apache-2.0; the MiniMax H3 weights, the wolf LoRA and fal's hosted use carry their own terms, so check those before client work.

Our take: this is the most practical new piece this week for game cinematics, creature previs and stylised 3D shorts. A roar or take-off cycle for about the price of a coffee changes what a small team can block out. It's the creature-shaped sibling of prompt2rig, which does the same video-model-as-actor trick for human rigs.