doubao-image — generate images with the local Doubao app

面向：使用者 · 状态：stable · 最近核对：2026-09-20

Registers one tool, generate_image, that drives the locally installed and logged-in Doubao (豆包) desktop app to produce a picture, saves the full-resolution file, and sends it into the conversation.

plugins/doubao-image/
├── package.json          # type: module, dsh.bundle.patch, exports
├── cordis.patch.yml      # bundle patch: the loader row
├── lib/
│   ├── index.js          # cordis plugin: the generate_image tool
│   ├── doubao.mjs        # app launch, prompt submission, result harvest
│   ├── cdp.mjs           # dependency-free Chrome DevTools Protocol client
│   ├── watermark.mjs     # removes Doubao's own overlay, with a self-check
│   ├── image.mjs         # minimal PNG codec + raster helpers
│   └── generate.mjs      # CLI driver, for testing without spending a model turn
└── test/                 # node --test (no DSH, no app, no network)

Install

# from npm
dsh plugin --profile web add dsh-plugin-doubao-image

# or from the repository root — the leading ./ matters: DSH anchors relative specs
# against your current directory, otherwise pnpm would resolve it inside the profile
dsh plugin --profile web add ./plugins/doubao-image

Then restart DSH (or reload the profile). The tool appears as generate_image.

Use

# exercise the whole path without going through the model
node plugins/doubao-image/lib/generate.mjs \
  --prompt "一只戴黄铜护目镜的橘猫，坐在双翼飞机驾驶舱里" \
  --style "黏土质感 C4D 渲染，暖色夕阳光" \
  --out ~/Pictures/doubao

The tool takes prompt (required) plus style, composition, aspect, negative, count, dir, and reuseCurrentChat.

How it works, and why it is built this way

Doubao ships a full Chromium (Contents/Helpers/Doubao Browser.app), and its launcher forwards unknown flags, so starting it with --remote-debugging-port yields a real CDP endpoint. That is the whole reason this works without the macOS Accessibility permission that osascript would need — which matters, because that permission is not granted on this machine.

The findings each step depends on, all established by probing the running app:

Step	What is actually true
Launch	Doubao holds a singleton lock. A normally-launched instance never opens the debug port, and launching again just forwards to it. The app must therefore be quit and restarted — the first call costs ~10–40 s.
Attach	doubao-chat is the real 1200×800 window. doubao-launcher is a 400×600 sidebar that also accepts prompts and renders results at 90 px, and doubao-chat is additionally served as an empty <html>hello</html> shell. Identification uses live window size plus the presence of the editor, never the URL alone.
Type	The box is div.tiptap.ProseMirror[contenteditable=true]. execCommand('insertText') goes through the same beforeinput path a keystroke does, so the app's own state and send button update. Assigning innerText would not.
Detect	The status sentence 「已生成 N 张候选图」 is the completion signal, not the image count — the grid mounts placeholders before the renders arrive.
Harvest	The picture in the grid is a cpreview_wm1 render behind an overlay, and its full-size sibling <img> is a bare data:image/svg+xml placeholder with no pixels, so the bytes cannot be read from the grid. Instead the tool clicks each thumbnail to open the image viewer and uses the app's own [data-testid=edit_image_download_button], which writes the file to ~/Downloads; the tool then moves it to dir.

dsh-doctor-style checks live in the unit tests: the output schema must stay in sync with execute, because a tool that returns a value without a declared schema is refused at registration and takes the whole plugin tree — and every DSH boot, the desktop app included — down with it.

The watermark

Doubao stamps 「豆包AI生成」 into the bottom-right of every saved picture. It is an overlay the app composites on, not part of the artwork, which is what makes removal a solved problem instead of an inpainting guess. All of this was measured, not assumed:

Property	Measurement
Position	Anchored to the bottom-right corner; a glyph mask from one image correlates with an unrelated image's corner at NCC 0.949 (inverse control −0.949, exactly symmetric)
Scale	Proportional to the image's short side, not its width: 2048²→321×69 box, 1365×2048→214×46 box, both giving width/short ≈ 0.1567. Verified on a second sample per shape
Blend	Plain alpha composite C = a·W + (1−a)·B, fitted at a = 0.638, W = 234.4, RMS residual 6.3/255
Soft edge	Glyph interiors measure a = 0.63 while the antialiased rim ramps to 0 — which is why a hard 0/1 mask left a visible outline

Removal inverts the composite per pixel, and estimates the artwork behind the mark by coarse-grid interpolation that excludes masked pixels. The output PNG is encoded losslessly (round-trip delta verified 0).

Self-check. The verifyOnFlat helper turns "does it look clean?" into a number: on a flat-background sample the artwork behind the mark is a constant, so any residual offset inside the glyph area is the watermark that survived. Measured on a flat 2048² sample: residual 46.3 → 0.0 (100% removed). On a photograph the same metric would be meaningless, so it detects a non-flat frame (background IQR > 8) and says so instead of printing a flattering zero.

Scope. This targets this one overlay, at this one anchor, with this one calibrated blend. It is not a general watermark remover. It exists for internal reference material — animation reference, storyboards, internal review. Doubao's official API (Volcengine Ark) returns images with no watermark at all, and that is the right channel for anything published or commercial.

Stripping is off by default (owner's decision, 2026-09-25): an image comes back with Doubao's own 「豆包AI生成」 badge unless the caller explicitly asks for stripWatermark: true. Keeping the badge is the safe default — it is what makes a picture traceable to the tool that made it — and the capability stays here for the internal-reference case described above.

Known limitations

• Generation consumes the user's Doubao quota (the app reports 「本次生成将 消耗每日免费额度」). Exactly one generation is issued per call; if the plugin cannot confirm a submission it stops rather than retrying, because a blind retry is what once produced duplicate generations.
• The app window is closed and reopened on the first call, and Doubao must stay running-and-instrumented for the harvest to complete.
• The harvest depends on Doubao's own UI markup (image-box-grid, edit_image_download_button). A Doubao redesign breaks it; the selector surface is deliberately kept to those two names.
• The watermark pass only re-encodes PNG. Other formats are saved exactly as downloaded rather than silently recompressed.