by bellisji233
Agent skill (Claude Code) that turns one character illustration into a half-body Live2D avatar for VTube Studio. 一张立绘做出 VTube Studio 半身 Live2D 皮套。
Available after the catalog security scan passes.
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM is an informational directory and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
The deep catalog scan for this skill is still queued. Run an instant dependency check now instead.
# Add to your Claude Code skills
git clone https://github.com/bellisji233/live2d-halfbody-skillGuides for using ai agents skills like live2d-halfbody-skill.
给不想手动建模、不想调网格的人。一张立绘或一段角色描述进去,Agent 一步步做出能在 VTS 里直播用的半身 Live2D 皮套。用户只做三件事:
通过 或 不通过:<现象>(第 5 步)。其余由 Agent 完成,并在交给用户前自查。
| 功能 | 驱动 |
|---|---|
| 转头左右、上下(±30),歪头时身体跟着倾斜 | 摄像头 |
| 身体跟随:左右转、前后、呼吸起伏 | 摄像头、VTS 自动呼吸 |
| 眨眼、单眼眨眼(快闭上时换成画好的闭眼);半闭眼时双眼皮淡出 | 摄像头(左右眼分开) |
| 眼珠跟随、眉毛上下 | 摄像头 |
| 张嘴说话、嘴角笑 | 摄像头 |
| 头发分段摆动:刘海(可选)、两侧(三段,从发根到发梢一段跟一段;可别在耳后)、后发或发髻;点头时上下弹。双马尾每条从发绳往下单独摆,同样分三段。过肩长发在肩膀以下跟身体走 | 物理 |
| 挂件(项链吊坠、耳坠)摆动 | 物理 |
| 眼镜跟着脸转 | 摄像头 |
| 表情:笑眼 N1、O 型嘴 N2、急眼 N3、困扰 N4,清除 N0 | 快捷键 |
| 手势:招手 N5、摊手 N6、指向 N7 | 快捷键(也配了摄像头手势,见“已知问题”) |
| 脸红 N8 | 快捷键 |
做不到:大角度转头(要侧脸素材)、全身、手臂跟着摄像头连续动(要分段手臂素材)。产出是 VTS 直接用的 .moc3,不是 Cubism Editor 的工程文件(.cmo3),Editor 打不开。精修级的效果(更细的物理、更多表情、商业级网格)要把 layers-st/ 的图层导进 Cubism Editor 重新绑定。
素材/角色名/ 立绘、提示词、变体、See-through 结果、图层
A0-base.png 定稿立绘
prompts/ 从 references/提示词/标准/ 复制,填好 [[…]]
check/ 每张变体的对位检查图
seethrough.psd See-through 的原始结果
seethrough/ 拆成每层一张 PNG
layers/ 自拆的图层:只作 See-through 合成的底,不单独交付
layers-st/ 最终图层(See-through 合成,含手势和脸红)
R0-rest.png 最终图层的静止合成,手势和脸红变体的底图
角色/角色名/
character.json 位置、幅度、图层对照
标注.png character.json 里的位置画在立绘上
素材检查/ prepared/ build/runtime/ 自查/ vts/
命令都在用户的工作目录(放 素材/、角色/ 的地方)里跑,命令前加 PYTHONPATH=<skill 目录>(如 ~/.claude/skills/live2d-halfbody),Python 才找得到 rigkit。
python -m rigkit.doctor
有 ✗ 的先处理完再出图,否则会在出完二十来张图之后才卡住。! 只影响一部分:没有 See-through 令牌时匿名额度不一定够,没有 ffmpeg 时出不了自查短片,没有 Codex 时用其他图像工具出图(见第 2 步)。
把 references/提示词/标准/ 整个复制到 素材/角色名/prompts/,先填好 A0-base.txt 里的 [[…]](外貌、刘海、两侧头发、头发长短、眼镜、耳坠)。
素材/角色名/参考.png,生成时作为参考图附上,得到一张按绑定要求画的版本(脖子露出;有刘海时停在眉毛以上;两侧头发盖住耳朵到下巴为止,或别在耳后;平光、白底、手臂和身体之间留缝)。原图不满足这些要求时,直接拆会在动起来时出问题。python -m rigkit.variants 素材/角色名 A0-base
没有 Codex 时按第 2 步的“用其他图像工具出图”出 A0。给用户看,用户确认后才往下做。
已有按 图层清单 分好层的文件时,跳到第 4 步。
把 prompts/ 里其余的 [[…]] 换成这个角色的描述;角色没有的部件删掉对应的提示词:帽子或发饰 V1,耳前的侧发 V2,刘海 V3,挂件 V9(S1 里同时删掉那一行,见 出图)。然后:
python -m rigkit.variants 素材/角色名
依次生成 13~17 张“整张图改一处”的变体,每张出对位检查图。偏移 4 px 以内、改动只在该改的地方算合格;不合格删掉重出,同一张最多三次。细节见 出图。
用其他图像工具出图(没有 Codex,或 Agent 自己能出图):variants 遇到已经有的图只做检查,不重出,所以可以先用别的工具把图出好放进去。
python -m rigkit.variants 素材/角色名 --list
每行一张图:要附的图、提示词文件、出了没有。按顺序一次出一张,附的图标着“还没有”时等它出好:
prompts/<名>.txt 的全文,不改写;只附列出的那一张图,让工具在它上面改。A0 附参考图,没有参考图时直接生成。素材/角色名/<名>.png,和附的图长宽比相同。不裁、不加边、不抠图、不修补。出完一批,跑一次 python -m rigkit.variants 素材/角色名 出检查图,合格标准同上。第 3 步的手势和脸红同样:--list F1-blush P1-wave P2-explain P3-point。还没用 Codex 以外的工具完整做过一个角色。
先用变体自拆,作为合成的底:
python -m rigkit.split 素材/角色名 素材/角色名/layers --pad-top 90
再问用户是否同意把立绘传到 See-through,同意后:
python -m rigkit.seethrough fetch 素材/角色名/A0-base.png 素材/角色名/seethrough
python -m rigkit.seethrough split 素材/角色名 素材/角色名/layers 素材/角色名/layers-st --pad-top 90
fetch 先试 Hugging Face,没跑成再换魔搭(--site hf 或 --site modelscope 指定一个),各用能给的最高分辨率。也可以把网页版、ComfyUI 或自己显卡跑出的 PSD 交给它:fetch 那个.psd 素材/角色名/seethrough。本机装了 See-through 时用 --site local:不传图,不占额度(见“环境”)。
最终图层必须走 See-through:自拆的图层动起来边缘有碎点、被挡住的头发补不全,不单独交付。当天没有额度时,换另一个网站、等第二天,或者用 PSD。
split 会逐条打印它对 See-through 图层做的修正(裁掉画到别处的、挪回归错的配饰、补漏画的部件),规则见 规则 的“See-through 合成”。看 layers-st/拆层总览.png:每层只有自己的东西,静止合成和立绘只差在画风。
闭眼、笑眼是自拆的。它们带进旁边的线或边缘毛躁、同一处改两次还在时,把这两张变体交给 See-through 拆,再跑一次上面的 split,它会改用 See-through 的睫毛层(规则第 103 条):
python -m rigkit.seethrough fetch 素材/角色名/D5-eyes-closed.png 素材/角色名/seethrough-D5-eyes-closed
python -m rigkit.seethrough fetch 素材/角色名/D6-eyes-smile.png 素材/角色名/seethrough-D6-eyes-smile
传到网站同样先问用户,每张占一次额度。本机跑时加 --site local --head-from 素材/角色名/seethrough,只拆头部。
然后做手势和脸红:
python -m rigkit.extras 素材/角色名 素材/角色名/layers-st # 写出 R0-rest.png
python -m rigkit.variants 素材/角色名 F1-blush P1-wave P2-explain P3-point
python -m rigkit.extras 素材/角色名 素材/角色名/layers-st # 加进图层
看 layers-st/手势和脸红.png:每行是静止、变体、模型里打开后的样子,手臂和身体接得上、手的边缘干净才往下做。
python -m rigkit.estimate 素材/角色名/layers-st 角色/角色名 --name 名字
python -m rigkit.check_art 角色/角色名/character.json
python -m rigkit.prepare 角色/角色名/character.json
python -m rigkit.build 角色/角色名/character.json # 几分钟
estimate 不覆盖已有的 character.json,写成 character.draft.json;图层改过要重估时,比对后换上。标注.png:脸的椭圆、眼眉鼻嘴、歪头转轴、颈根、肩、腰、发根线不在对的位置时,改 character.json 里对应的数字,重新 prepare 和 build。素材检查/:露底 有品红、原画对照 有成线的红色时,回到第 3 步。透出背景 列出静止时没盖住的地方、手势时变透明的地方,再看 青底.png(静止和各手势放在青色背景上):白衣服、浅色皮肤处的缺口在白底上看不出。手臂挪开处变透明是正常的,其余回到第 3 步。自查:
python -m rigkit.preview flips 角色/角色名/build/runtime
python -m rigkit.preview sheet 角色/角色名/character.json 角色/角色名/自查 新版=角色/角色名/build/runtime
python -m rigkit.preview clip 角色/角色名/character.json 角色/角色名/自查/动作.mp4 新版=角色/角色名/build/runtime
翻折三角形应为 0。看放大图和短片里的固定检查区(眉毛和刘海、脸边、下巴和脖子、太阳穴和耳朵、眼角和睫毛、嘴角),有问题先修,再交给用户。有上一版时把上一版放在前面并排比。
再在 VTS 的背景色上,把闭眼、半闭、笑眼、张大嘴、半张嘴、三个手势、左右甩头各渲一张,眼睛、嘴、头发交界放大三到四倍看:用户在 VTS 里最先看到的就是这些画面。
python -m rigkit.vts 角色/角色名/character.json --install
新模型可以在 VTS 开着时装;覆盖已装的模型要先退出 VTS,这时保留用户在 VTS 里调过的设置,只补缺的快捷键和映射。VTS 开着又要换新版时,改 vts.name 和 vts.folder 装成新文件夹。
给用户并排图、短片和下面这张清单,回复 通过 或 不通过:<现象>:
| 看什么 | 怎么做 |
|---|---|
| 转头、歪头 | 左右上下转头、歪头:脸、头发、帽子、耳朵跟着转,没有裂缝、碎点、留在原地的东西 |
| 眼睛 | 眨眼、单眼眨眼、半闭眼;转头时睫毛不穿过头发 |
| 嘴 | 说话、笑;闭着嘴转头时不歪嘴 |
| 头发 | 甩头:刘海、侧发、后发分段摆,从发根到发梢一段比一段晚;点头时上下弹;长发在肩膀以下跟身体 |
| 挂件 | 身体动时吊坠摆动,底下的衣服完整 |
| 身体 | 身体跟随左右转、前后,呼吸起伏 |
| 表情 | N1~N4 切换,N0 清除 |
| 手势 | N5~N7:身体画风不跳,手的边缘干净;手势打开时转头,手臂跟着身体走 |
| 脸红 | N8:边缘柔和;甩头时头发下面不露边 |
动作幅度都在 character.json:
| 想改 | 字段 |
|---|---|
| 歪头角度 | head.tilt_deg(满幅时的度数) |
| 转头时脸走多远 | head.yaw_px、head.pitch_px |
| 身体倾斜 | body.tilt_deg(每单位参数的度数,满幅 ×10) |
| 呼吸起伏 | body.breath.head、body.breath.shoulder(像素) |
| 身体跟随 | body.turn、body.shift、body.lean_y |
| 头发摆幅 | hair.locks.<图层>.swing;中段和发梢 hair.locks.<图层>.bend;点头 hair.locks.<图层>.pitch |
| 物理手感 | hair.locks.<图层>.chain(三段)或 physics:delay 越小越拖、length 越长越慢 |
| 挂件摆幅 | dangles.<图层>.deg(满幅时的度数) |
| 半闭眼 | eyes.crease_fade(双眼皮淡出的区间)、eyes.lower_from(下眼睑开始合的位置) |
| 嘴从哪里开始张 | mouth.open_from:摄像头读数低于它时保持闭嘴 |
幅度按 VTS 半身取景和真人动作判断:正常歪头约 15°,模型只转 3° 就等于没有。改完重新 build。其余字段见 配置说明。
全部规则和来由见 规则。最常被违反的几条:
pip install -r <skill 目录>/requirements.txt(系统 Python 不让装时建一个虚拟环境)。出短片要 ffmpeg。codex(在 PATH 里即可),登录 ChatGPT 账号;都不是时把环境变量 CODEX 指向它。没有 Codex 时用其他能改图的图像工具,见第 2 步。HF_TOKEN 或 ~/.cache/huggingface/token(hf auth login 存的位置),每天约 1~2 次。MODELSCOPE_API_TOKEN 或 ~/.cache/modelscope/token。要排队,一次可能半小时。1536 的头发轮廓和眼睛比 1280 更接近原画。SEETHROUGH_DIR 指向它,装法见 出图 的“本机跑 See-through”。CUBISM_CORE 指向其中的 Core 动态库。Core 按 Live2D 的许可使用,不随本 skill 分发。macOS 上验证过;Windows 未验证。
See how live2d-halfbody-skill compares with popular alternatives.
English | 简体中文
Give an AI agent (such as Claude Code) one character illustration or a text description, and it builds a half-body Live2D avatar for streaming in VTube Studio, step by step. No manual modeling, no Cubism Editor, no mesh editing.



Not supported: large head turns, full body, arms following the camera continuously.
The output is a model file VTube Studio loads directly (.moc3), not a Cubism Editor project, so the Editor can't open it. For production-grade results (finer physics, more expressions), import the separated layers into Cubism Editor and rig them again. Suited to front-facing half-body characters with short hair, hair past the shoulders or twin tails.
The agent redraws your illustration or description into a version that separates cleanly into layers, and you approve it before the next step. The face, hair color, eyes, clothing, palette and art style stay the same; these are adjusted:
For hairstyles or accessories outside this list (animal ears, braids), ask the agent to extend the skill: add drawing rules, adjust the layer-splitting and rigging code, then check the result in VTube Studio.
HF_TOKEN environment variable or ~/.cache/huggingface/token. Without logging in, the daily free quota may not cover one run; logged in, about 1–2 runs a dayMODELSCOPE_API_TOKEN or ~/.cache/modelscope/tokenTested on macOS. Not tested on Windows.
git clone https://github.com/bellisji233/live2d-halfbody-skill ~/.claude/skills/live2d-halfbodypip install -r requirements.txtOne character takes 18–22 generated images and one See-through run; from illustration to VTube Studio is about an hour, see FAQ.
Timings from the third character (glasses):
| Step | Time | Who |
|---|---|---|
| 1. Generate the illustration (Codex) | about 1.5 min | Agent |
| 2. Approve the illustration | — | You |
| 3. Generate 13–17 variants: eyes closed, mouth open, hair removed, etc. (Codex, about 1.5 min each) | about 20–30 min | Agent |
| 4. See-through layer separation (Hugging Face, runs alongside step 3) | about 5 min | Agent, asks you before uploading |
| 5. Split layers and merge with the See-through result | about 2 min | Agent |
| 6. Blush and three gestures: 4 more images | about 7 min | Agent |
| 7. Estimate positions, check the art, build the model | about 1 min | Agent |
| 8. Self-check, install into VTube Studio | 1–2 min | Agent |
| 9. Check the avatar in VTube Studio | 5–10 min | You |
Machine time adds up to about 30–45 minutes; with your approval and check, a character can be in VTube Studio within an hour if nothing goes wrong. The number of variants depends on whether the character has a hat or hair accessory, bangs, side hair in front of the ears, and pendants.
When the Codex or See-through quota runs out, the run stops at that step. Run it again once the quota resets; images already generated are not redone.
The skill instructions (SKILL.md, references/) and script messages are in Chinese. The agent reads them as they are; talk to it in any language.
MIT, see LICENSE. rigkit/vendor/py-moc3 is MIT, license included. Live2D Cubism Core is not distributed with this skill; it is used under Live2D's license.
live2d-halfbody-skill is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by bellisji233. Agent skill (Claude Code) that turns one character illustration into a half-body Live2D avatar for VTube Studio. 一张立绘做出 VTube Studio 半身 Live2D 皮套。. It has 50 GitHub stars.
live2d-halfbody-skill's catalog security scan is still queued. You can run an instant dependency and prompt-injection check now with the "Scan for vulnerabilities" button above.
Clone the repository with "git clone https://github.com/bellisji233/live2d-halfbody-skill" and add it to your Claude Code skills directory (see the Installation section above). live2d-halfbody-skill ships a SKILL.md manifest, so compatible agents can discover and load it automatically.
live2d-halfbody-skill is primarily written in Python. It is open-source under bellisji233 on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh live2d-halfbody-skill against similar tools.
No comments yet. Be the first to share your thoughts!