anymove (creature motion capture from one AI video)
/anymove
Patch's open-source creature mocap: a video model films a rigged 3D creature from two cameras in one frame, a DINOv2 joint reader trained on that creature finds every joint in both views, and the rig is fitted to both views and exported as a GLB. Six agent skills plus the scripts.
- 许可证
- Apache-2.0
- 出品
- GameSkill 出品
我们测过
在 Blender 5.2.3 上通过 23/23 项检查（2026-10-29）：能触发、有可运行示例、有截图和自检清单、单一职责、中英文文档齐全。触发检查是把提示词和 skill 描述做匹配，不是真实 agent 运行。
测试报告这个 skill 做什么
什么时候触发
Motion capture for any creature from one AI video, with the anymove repo. …
GameSkill 测过的版本
GameSkill 调优版：skill 文件取自提交 7a82a0a8fadf，按我们的 skill 流程加上可运行示例、测试报告、自检清单和中文指南。
协议：基于 patch-gg/anymove（作者 Patch），协议 Apache-2.0。ZIP 保留协议文件（LICENSE），CHANGES.md 列出了 GameSkill 加的所有文件。
自检清单
- 先在 CPU 上跑一遍自带的龙吼拟合，确认环境没问题 / Run the bundled dragon-roar fit on the CPU first to check the setup
- 新生物先摆骨架、给每条腿涂不同颜色，再拍片 / Place the skeleton and give every limb its own colour before filming a new creature
- 拍片和 LoRA 用你自己的 fal key，花钱之前先看报价 / Clips and LoRAs use your own fal key; check the price before spending
- 交付前看两视角一致分数和重投影误差，过不了就重拍 / Check paw agreement and reprojection error before delivery; refilm when they fail
- 八条腿走路、趴下、打滚目前会失败，别硬拟合 / Eight-legged walks, lying down and rolling over fail today; do not force a fit
- 示例网格、视频和 LoRA 不是 Apache-2.0，商用前看 Tripo、MiniMax 和 fal 的条款 / The example meshes, clips and LoRA are not Apache-2.0; read Tripo, MiniMax and fal terms before commercial use
推荐搭配
中文指南
这个 skill 做什么
给任意生物做动作捕捉，只要一段 AI 视频。做法是在一个画面里放两台正交摄像机：左边正面、右边侧面，比例相同、共用竖轴，同一个关节在两边的高度一样。视频模型（作者用 fal 上的 MiniMax H3 Max）照着提示词让生物在两个视角里同时动；再用这只生物的随机姿势训练一个 DINOv2 关节识别器，找出两个视角里的每个关节；最后让骨架同时贴合两个视角，导出 GLB。双机位的思路来自 @CanItCode，LoRA 那一步来自 @Graalitoo。
包里是 6 个 skill：anymove（总入口，告诉 agent 下一步该做什么）、anymove-rig（网格、骨架、蒙皮、给四肢上色）、anymove-film（写提示词、拍双机位视频）、anymove-reader（训练关节识别器）、anymove-capture（读关节、拟合、导出 GLB）、anymove-lora（训练让版式更稳的 H3 LoRA），以及 upstream-support/ 里的全部脚本。
什么时候用
- 给龙、蜘蛛、狼、怪物这类非人形生物做游戏动画，没有动捕设备
- 想从一句话或一段 AI 视频得到自己骨架上的 GLB 动画
- 人形角色用 Mixamo 骨架的话，看 Mixamo LLM Mocap 更合适
怎么装
Claude Code 里执行 /plugin marketplace add Patch-gg/anymove；或者下载本页 ZIP，把 anymove-creature-mocap/ 里的 6 个 skill 文件夹放进 .claude/skills/ 或 .agents/skills/。upstream-support/ 是仓库的脚本（mocap/、blender/、tools/、assets/、prompts/、examples/），skill 要求在这份目录里运行，没有的话它会让 agent 去 clone 原仓库。环境：Python 3.11、Blender 5.x、pip install -r requirements.txt。训练识别器、读新视频要 NVIDIA 显卡（训练 24 GB）；拍片和 LoRA 要你自己的 fal key（README 写的价格：网格约 $0.40，5 秒视频约 $0.26，LoRA 约 $12）。没有显卡可以装 CPU 版 torch，只能跑最后的拟合。
最小示例
gameskill/examples/tests/run_anymove_hands_on.py 不用显卡、不用 fal、不下载任何东西：用仓库自带的龙吼关节数据（examples/dragon_roar/det.npz）在 CPU 上跑 mocap.solve，检查动作数据；再用我们写的 standin_rig.py 按 rig_info.json 重建骨架、配一个每根骨头一个方块的替身身体（不用 release 里的 Tripo 网格），交给仓库自己的 blender/apply_motion.py 打关键帧、渲染、导出 GLB，然后检查 GLB。本页截图是替身身体在第 0、60、110 帧的侧面和正面。
我们怎么测的
总分 23/23。运行部分 11 项全部通过：CPU 上跑 mocap.solve 约 30 秒（README 说 20 秒；不限线程、和别的任务抢 CPU 时我们测到过 8 分钟，所以脚本里设了 OMP_NUM_THREADS=4）；两视角一致（重投影误差正面 5.93 像素、侧面 7.99 像素，爪子高度一致分数 0.93）；124 帧 × 42 根骨头都有数据，数值都有限、根骨四元数是单位长度；15252 个欧拉角没有一个超出 rig 的关节限制；我们按 rig_info.json 重建的骨架和原 rest 矩阵误差 7e-7；仓库自己的 apply_motion.py 打完关键帧后，Blender 算出的关节位置和拟合结果最多差 0.1 毫米（龙高 1.84 米）；GLB 里 1 个 skin 42 个关节，42 条旋转通道每条 124 个关键帧，时长 5.12 秒；侧面和正面各渲染 3 帧。触发测试：3 句正向提示词都排第 1，2 句反向提示词排第 5、第 6。6 个 skill 的 description 都写了什么时候用；anymove-reader 写的是 “Use after anymove-rig and before anymove-capture, when …”，我们的检查原来只认 “Use when/for/to”，这次把 “Use after/before” 也算进去了（改的是我们的检查脚本，没改作者的文件）。没测的：做网格、摆骨架和蒙皮新生物、拍双机位视频、训练识别器、读新视频、LoRA，这些要 NVIDIA 显卡或 fal 付费。ZIP 里作者的文件一个没改。结果在 gameskill/test-report.json。
授权和不在 ZIP 里的东西
代码和 skill 是 Apache-2.0，ZIP 带原样的 LICENSE。Patch 的名称、标志、字标和 Patch Pixel 字体不在授权里，所以 README 里的 Patch 标头图、其他公司的标志（docs/readme/）、演示 GIF 和步骤图（docs/anymove.gif、docs/steps/，画面是 Tripo 网格和 MiniMax 生成的视频）都没放进 ZIP，README 里这些图片会显示不出来。release 里的示例网格（Tripo 生成）、视频（MiniMax H3 生成）适用 fal 和各自的条款，狼的 LoRA 适用 MiniMax H3 社区许可证；tools/fetch_examples.sh 会下载它们，用之前先看条款。
搭配
ai-mesh-weapon-rig：AI 网格拆件、刚体绑定、开火换弹动画blender-export-pipelines：把 GLB 稳定导进引擎threejs-glb-scene-assembly：在网页里摆放、播放 GLB
安装
告诉你的 AI
Install the skill from https://gameskill.dev/skills/anymove-creature-mocap/install.md
GameSkill 测过的版本
mkdir -p .claude/skills
curl -fsSL https://gameskill.dev/dl/skills/anymove-creature-mocap.zip -o /tmp/gameskill-anymove-creature-mocap.zip
unzip -oq /tmp/gameskill-anymove-creature-mocap.zip -d .claude/skills
rm /tmp/gameskill-anymove-creature-mocap.zip或者手动安装
git clone --depth 1 https://github.com/patch-gg/anymove.git /tmp/gameskill-patch-gg-anymove
rm -rf /tmp/gameskill-patch-gg-anymove/.git
mkdir -p .claude/skills
cp -R /tmp/gameskill-patch-gg-anymove/skills/anymove/ .claude/skills/anymove
cp -R /tmp/gameskill-patch-gg-anymove/skills/anymove-rig/ .claude/skills/anymove-rig
cp -R /tmp/gameskill-patch-gg-anymove/skills/anymove-film/ .claude/skills/anymove-film
cp -R /tmp/gameskill-patch-gg-anymove/skills/anymove-reader/ .claude/skills/anymove-reader
cp -R /tmp/gameskill-patch-gg-anymove/skills/anymove-capture/ .claude/skills/anymove-capture
cp -R /tmp/gameskill-patch-gg-anymove/skills/anymove-lora/ .claude/skills/anymove-lora
rm -rf /tmp/gameskill-patch-gg-anymove完整仓库安装
/plugin marketplace add Patch-gg/anymovegit 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 # add your own FAL_KEYWithout an NVIDIA GPU, install the CPU build of torch (--index-url https://download.pytorch.org/whl/cpu): the fitting step runs, training the reader does not.
命令照搬项目 README。在本机运行前，请先看一眼仓库。
包含内容
- A router skill (anymove) and five step skills: rig, film, reader, capture, LoRA
- mocap/: the DINOv2 joint reader, training, and the two-view rig fit in torch
- blender/: build and skin the rig, paint the limbs, render datasets, apply a motion and export a GLB
- Rigs and prompts for four example creatures (dragon, giant spider, fire demon, wolf) and the dragon's roar as read by its reader, to try the fit on a CPU
适合
- Animating non-human creatures (quadrupeds, dragons, monsters) for a game without a mocap suit
- Turning a text prompt or an AI clip into a GLB animation on your own rig
不适合
- Machines without an NVIDIA GPU, beyond the CPU fit of the bundled example
- Eight-legged walks, lying down or rolling over (the README lists these as failing)
- Humans on Mixamo rigs: see Mixamo LLM Mocap
关于这个 Skill
anymove is Patch's motion capture for any creature from one AI video. One image holds two orthographic views of the rigged creature, front on the left and side on the right, sharing the vertical axis, so a joint sits on the same pixel row in both halves. A video model animates both views at once from a prompt, a DINOv2 joint reader trained on that one creature finds every joint in both views, and the rig is fitted to both at once and exported as a GLB.
The repo ships six agent skills (a router plus rig, film, reader, capture and LoRA steps) and the scripts they run, with rigs and prompts for a dragon, a giant spider, a fire demon and a wolf. Training a reader and reading new clips need an NVIDIA GPU; clips and LoRAs run on fal with your own key. The two-camera method is credited to @CanItCode and the LoRA step to @Graalitoo.
Related
- Mixamo LLM Mocap for human motion on Mixamo rigs.
- MCP for Blender to drive Blender live from an agent.
环境要求
- Python 3.11 and Blender 5.x
- An NVIDIA GPU with CUDA to train a creature's joint reader (24 GB at batch 20) and read clips
- Your own fal.ai key for meshes and clips (about $0.40 a mesh, $0.26 a 5 s clip, about $12 a LoRA, per the README)
相关主题
常见问题
基于 patch-gg/anymove （作者 Patch，协议 Apache-2.0）。ZIP 里保留原作者的 LICENSE 和 NOTICE，ATTRIBUTION.md 和 CHANGES.md 写明来源提交和 GameSkill 加的内容。
搭配使用
作者标注的相关 Skills，或更适合相邻任务的 Skills。
Mixamo LLM Mocap
Turn a locked-camera video, filmed or AI-generated, into a clean FK animation on any Mixamo character in Blender. GVHMR pose estimation, SMPL-X retargeting, foot locks and fidelity checks, built to be run end to end by an AI agent.
MCP for Blender (blender-mcp)
Community plugin to control Blender with any LLM. The most-starred game-adjacent MCP server on GitHub.
Blender Animation
Animation in Blender 5.2: layered actions, Python keyframing, NLA, mocap retargeting, baking rigs to export skeletons and per-clip export.
相似 skills
Scenario Skills
Production-ready images, video, audio and 3D from any agent through the Scenario MCP, plus expert teams for Blender, Unreal and Unity.
Blender Hard-Surface Modeling (Scenario)
Hard-surface modeling in Blender for props, weapons and vehicles: booleans, bevels, weighted normals and support loops, plus shading fixes.
Blender Rigging with Python (Scenario)
Rig characters, creatures and props in Blender through Python: Rigify, weights, IK/FK, drivers, corrective and facial shape keys, and deformation tests.
Charakuru (きゃらくる)
A Tripo-sponsored kit that turns one full-body character illustration into a VRM 1.0 avatar with an AI agent doing the Tripo, Blender and rigging work, including hair spring bones and expressions. Source-available under its own terms, not open source, so see the repository for use.