---
type: "post"
id: "patch-releases-anymove-open-source-creature-motion-capture-d8d368"
slug: "patch-releases-anymove-open-source-creature-motion-capture-d8d368"
title: "Patch releases AnyMove open-source creature motion capture from one AI video"
description: "Patch released AnyMove, an open-source pipeline that turns AI-generated video of a creature into a rigged 3D animation exported as GLB."
retrieval_nugget: "Patch released AnyMove, an open-source pipeline that turns AI-generated video of a creature into a rigged 3D animation exported as GLB. Its two-view reconstruction approach targets bodies beyond human proportions, while published cost and quality figures make it possible for animation teams to judge whether the workflow is worth testing."
published_at: "2026-11-07"
updated_at: "2026-11-07"
record_date: "2026-11-07"
date_kind: "published_at"
topics: ["generative-media","open-source"]
entities: ["Patch","AnyMove","DINOv2","MiniMax H3 Max","fal","Blender"]
related_patterns: ["generative-media-becomes-application-infrastructure"]
novelty_label: "new_fact"
what_changed: "Patch releases AnyMove, an open-source pipeline that captures creature motion from one AI video and outputs a GLB rig using two camera views and DINOv2."
prior_refs: []
source_url: "https://github.com/patch-gg/anymove"
source_urls: ["https://github.com/patch-gg/anymove"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:patch-releases-anymove-open-source-creature-motion-capture-d8d368"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/patch-releases-anymove-open-source-creature-motion-capture-d8d368/","markdown":"https://newruntime.com/posts/patch-releases-anymove-open-source-creature-motion-capture-d8d368.md","json":"https://newruntime.com/posts/patch-releases-anymove-open-source-creature-motion-capture-d8d368.json"}
markdown_url: "https://newruntime.com/posts/patch-releases-anymove-open-source-creature-motion-capture-d8d368.md"
---

# Patch releases AnyMove open-source creature motion capture from one AI video

## Retrieval answer

Patch released AnyMove, an open-source pipeline that turns AI-generated video of a creature into a rigged 3D animation exported as GLB. Its two-view reconstruction approach targets bodies beyond human proportions, while published cost and quality figures make it possible for animation teams to judge whether the workflow is worth testing.

## Summary

Patch released AnyMove, an open-source pipeline that turns AI-generated video of a creature into a rigged 3D animation exported as GLB. Its two-view reconstruction approach targets bodies beyond human proportions, while published cost and quality figures make it possible for animation teams to judge whether the workflow is worth testing.

## What happened

Patch’s GitHub repository describes AnyMove as a working implementation of a two-camera motion-capture method. A video model generates a clip showing the creature from front and side views simultaneously in one frame. A DINOv2-based reader, trained on that creature, locates joints in both views. The pipeline intersects the two views to estimate 3D joint positions, fits a rig to those positions frame by frame, and exports an animation as GLB. The README says the output is 24 frames per second, with no hand edits in the fitting step; reading a clip takes about 10 seconds and fitting about 45 seconds.

The workflow is creature-specific. A builder starts with a mesh, places the skeleton joints on grid views, and uses Blender to build and skin the rig. The project then renders about 2,200 random poses to train a joint reader for that creature. New video clips are generated through fal, using MiniMax H3 Max in the examples. The repository includes Claude Code agent skills to guide setup, rigging, clip generation, capture, quality checks and packaging. It also includes example assets and a sample dragon capture that can be fitted without repeating every generation step.

The README reports a mean reader miss falling from 40 pixels to 6 pixels for its training setup. In a table of example creatures, it reports 0.94 cm mean 3D joint error for a wolf exam; scores for dragon motions range from 0.81 for walking to 1.00 for takeoff. These are project-reported tests, not an independent benchmark. A separate run generated 335 clips, of which 193 passed checks, or 58%. A usable move took about seven clips on average. The repository estimates about $3 per usable move, including generation and rented compute, and about $17 for a new creature with roughly five moves. It estimates about $11 for the latter when using the builder’s own GPU.

The project is released under Apache-2.0, but that does not automatically license generated assets or every model used in the workflow. Patch says the MiniMax H3 weights carry a community license and hosted use on fal is subject to fal’s terms; users should check both before shipping. The README specifies an NVIDIA GPU with 12 GB memory for training a reader, with 24 GB described as safer, plus Python 3.11, Blender 5.x and a fal key to make new clips. There is no cloud version. The optional H3 LoRA step is estimated at about $12 and two hours on fal.

## What the archive adds

New Runtime’s earlier coverage of Claude Motion on 2026-10-12 described an animation workflow built for coding agents. AnyMove is a new event, with a different emphasis: turning generated creature footage into skeletal animation for non-human body plans. The earlier note is not linked here because no public New Runtime URL was supplied for it.

## Trend signal

This supports [Generative media becomes application infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/): AnyMove treats generated video as an input to a repeatable asset pipeline, not merely as a finished artifact. Its output is a reusable rigged animation file, though the creature-specific setup and manual rig placement show that this is not yet a broadly automated production surface.

## Why now

Patch’s repository presents the project as a working pipeline and includes cost figures from a run performed in October 2026. Those numbers give teams a practical basis for evaluating the workflow now, while also making clear that repeated video generation and per-creature training remain part of the cost and effort.

## Why it matters

For game and animation teams, the notable shift is the proposed handoff between generative video and a conventional production asset: a GLB with motion on a real skeleton. If the method works for a specific creature, teams can prototype a roar, slam, takeoff or other large gesture without relying on human-centric motion capture assumptions. The per-move estimate is also more useful than a per-clip price because most clips fail the project’s checks.

The trade-off is substantial setup and validation. Each new creature needs its own reader, joint layout and rig, while poor takes still consume generation budget. The README reports foot sliding of roughly 1–2 cm per step in examples and says eight-legged walking, tail-tip tracking and fine contact are weak areas. A team should therefore evaluate it on a constrained use case—such as a distinctive one-off creature gesture—and compare the resulting animation with its existing rigging and mocap workflow before treating the economics as repeatable.

## Limits and uncertainty

The performance and cost figures are reported by Patch, and the README does not establish how they generalize beyond the listed creatures and motions. Its own tests use AI-derived video and, for the reader exam, made-up motions rendered in the same setup; these do not demonstrate accuracy against independent ground-truth motion capture. Video-model drift, view disagreement, foot sliding and failures on some body plans are documented limitations. Costs also depend on retries, GPU rental, model choice and time spent placing skeletons, for which the project does not provide a labor estimate. The repository’s license does not settle rights in outputs made with third-party models or hosted services.

## Skill to build now

A useful first test is to run the included dragon example through fitting and inspect the resulting GLB against its source clip. Then score one simple, large gesture on a creature close to your intended body plan. Record failed takes as well as successful ones: the project’s own average of about seven clips per usable move shows why a single showcase result is not a production estimate.

## Reader takeaway

AnyMove offers a practical experiment in converting AI-generated creature footage into reusable skeletal animation, with transparent enough numbers to estimate a pilot. Treat its cost and quality figures as project-specific, and test the exact body plan and motion you need before building it into a production pipeline.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Generative Media](https://newruntime.com/topics/generative-media.md): Explore the generative-media topic hub.
- topic hub: [Open Source](https://newruntime.com/topics/open-source.md): Explore the open-source topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)