lemo-wake

把你的照片叫醒 Wake your image.

一张图进去，一段约 5 秒、无缝循环的动图出来。 One picture in, a seamless loop of about five seconds out.

▶ 去作品墙看看 See the gallery · ▶ Logo 动效 Logo Motion

English | 中文

中文用户请看这里：完整中文说明 README.zh-CN.md English readers: keep reading below.

What it does 它做什么

Give Claude Code one picture and say "make it move". You get a loop of about 5 seconds whose end joins its start seamlessly. Photos, portraits, posters, illustrations, logos, charts, stickers, Xiaohongshu covers and chat screenshots all work.

It is not a filter. Claude first reads the image, takes it apart into text, shapes, objects, people and light, rebuilds the pieces in code and choreographs a short film for this one picture. It then renders the frames in a headless browser, joins the end back to the start and encodes the result.

• No image generation. Everything comes from your picture, from code and from small local models.
• No API keys. No cloud services. Everything runs on your machine.
• Your size, your formats. It asks once: keep your image's size or pick another ratio, and which formats you want.

中文简介

把一张图交给 Claude Code，说一句“让它动起来”，就能拿到一段约 5 秒、首尾无缝相接的动图。照片、人像、海报、插画、Logo、图表、表情包、小红书封面、聊天截图都可以。

这不是套滤镜。Claude 会先读懂画面，把文字、图形、物体、人物和光拆开，用代码重建，再为这张图编排一段专属的动作。最后在无头浏览器里逐帧渲染，把结尾接回开头，编码出片。

• 不生图：素材只来自你的图、代码和本地小模型。
• 不用 API key：不连云服务，全部在你的电脑上完成。
• 尺寸和格式你来选：开始前问一次，保持原图尺寸还是换个比例，要哪几种格式。

How it works 流程

Install

Built and tuned for Claude Code. Other agents (Codex, for example) can read the skill too, but the results are not guaranteed.

As a Claude Code plugin (recommended). In a terminal (both lines can be pasted at once):

claude plugin marketplace add lemomo-ai/lemo-wake
claude plugin install lemo-wake@lemo-wake

Or inside Claude Code, one line at a time (run the first, then type the second):

/plugin marketplace add lemomo-ai/lemo-wake
/plugin install lemo-wake@lemo-wake

Then restart Claude Code, or run /reload-plugins.

As a plain skill (without the upload and layers commands):

git clone --depth 1 https://github.com/lemomo-ai/lemo-wake
cp -R lemo-wake/skills/wake ~/.claude/skills/lemo-wake

You can also put it in a project's .claude/skills/lemo-wake.

Use

Give Claude an image and say "make it move". No command needed. Or use the commands:

With a plain skill install, the command is /lemo-wake <image>.

Layers. /lemo-wake:layers, or just "split this into layers" / "give me a PSD": each person, object and text block on its own transparent layer, the background behind them filled in, plus a depth map. You get layers.psd (Photoshop, Affinity, Procreate, or free in the browser at Photopea), the PNG layers and a layers.json with positions and stacking order (After Effects, Figma, game engines). People are kept whole, text stays pixels, and large hidden areas come out soft, since nothing is generated. It also reuses the parts of a film you already made.

How cleanly a picture splits depends on the picture, so the result is not guaranteed. If something is off, just tell the agent what you want, for example "put the two people on separate layers", "keep the flowers with the vase" or "clean up the edge of the hair", and it will redo that part.

What you get

Before it starts, Claude asks once: which size (your image's own size by default, or another ratio such as 16:9, 9:16 or 1:1) and which formats. Say "go" to take the defaults. Every format comes from the same frames, and more can be added later without re-rendering.

Why several: most chat apps and phone galleries don't play animated WebP, social platforms want MP4 for posts, and GIF works everywhere but is heavy. If a platform has another limit, such as "GIF under 2 MB", tell Claude.

The Live Photo is a .pvt bundle (Finder shows it as one file). AirDrop that one file to your iPhone and it lands in Photos as a Live Photo. Sending the photo and the video inside it separately gives you a photo and a video.

Photos with a long side over 2560 px are rendered at 2560 to keep it fast. Ask for full resolution if you need it.

First run

On first use Claude runs node scripts/doctor.mjs in the skill folder (skills/wake/). With --fix it installs what is missing.

Tested on macOS. Linux and Windows should work. If a step fails, Claude works out what is missing.

Local models (optional)

Small models that run locally on the CPU cut out objects and people, fill the space behind them and estimate depth. Most films get better with them: that is what lets things leave their place and really move. Nothing is downloaded until you say yes once, usually in the first question before a film starts.

About 600 MB in all. Every file is checked by sha256.

Asked once. Before the first film that needs a model, Claude tells you which one, what it is for, how big it is, where it comes from and where it will be stored, and asks once. Your answer is saved. If you say no, Claude finds a compromise in code and still finishes the film.

Where they are stored. With a plugin install, in ~/.claude/plugins/data/<plugin-id>/models/, kept across updates and removed on uninstall. With a plain skill install, in models/ in the skill folder. Set LEMO_WAKE_MODELS=<folder> to use another folder.

Commands. Run them in skills/wake/, or just ask Claude:

scripts/models.py                     # status
scripts/models.py fetch all --yes     # download everything now
scripts/models.py consent yes|no      # change your answer

Mainland China

Hugging Face is often unreachable from mainland China. Use the mirror. Files are still checked against the same sha256:

HF_ENDPOINT=https://hf-mirror.com scripts/models.py fetch all --yes

To use the mirror for every download, add it to ~/.claude/settings.json:

{ "env": { "HF_ENDPOINT": "https://hf-mirror.com" } }

If a download fails, Claude also tries a mirror on its own. For other slow downloads:

• Python packages: UV_DEFAULT_INDEX=https://pypi.tuna.tsinghua.edu.cn/simple
• npm: npm config set registry https://registry.npmmirror.com
• Headless browser: an installed Google Chrome works too, with CHROME=<path>
• Models by hand: download them from the links above, rename them to depth_anything_v2_small.onnx, birefnet_lite.onnx, modnet.onnx, mobile_sam_image_encoder.onnx, sam_mask_decoder_multi.onnx and lama_fp32.onnx, and put them in the models folder

Community

The gallery is not only for our samples. You are welcome to add your own films. Community films sit on the same wall as the samples, with an orange "Community" badge in the corner of the card and the author's name. The gallery also has a "Community" filter that shows only shared films.

How to share

1. Make a film you like with