Use when generating images through the Banana Studio backend with pw authentication. Supports prompt-only usage, storyboard file workflows with tool-rendered captions, and frontend-compatible modelId, imageOptions, layoutGuideImage, and referenceImages.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Use when generating images through the Banana Studio backend with pw authentication. Supports prompt-only usage, storyboard file workflows with tool-rendered captions, and frontend-compatible modelId, imageOptions, layoutGuideImage, and referenceImages.
Banana Studio Generate
Use this skill when you need to generate images through the Banana Studio backend instead of calling a model provider directly.
What it does
Sends pw directly on backend requests
Uses the same direct pw auth path as the Studio frontend, via x-banana-pw
Calls POST /api/generate/stream with the same request shape as the frontend
Supports prompt-only runs
Supports storyboard-file runs that generate each cell separately and render captions with the tool via POST /api/export/professional-preview
Storyboard layouts are customizable; the bundled markdown files are examples, not fixed presets
In storyboard mode, the script automatically chooses the closest supported image aspect ratio for each cell based on canvas size and row/column layout, which reduces export-time cropping
Supports the same request fields the frontend sends:
prompt
modelId
imageOptions.aspectRatio
imageOptions.imageSize
imageOptions.layoutRows
imageOptions.layoutColumns
layoutGuideImage
referenceImages
Supports reference images from either local file paths or remote http(s) image URLs
Default workflow
Run the bundled script at scripts/banana_studio_generate.mjs:
node /absolute/path/to/banana_studio_generate.mjs --prompt "A banana astronaut on the moon"
For captioned storyboard exports, pass --storyboard-file:
node /absolute/path/to/banana_studio_generate.mjs \
--prompt "A banana astronaut on the moon" \
--pw "test-banana"
Pass frontend-style options directly:
node /absolute/path/to/banana_studio_generate.mjs \
--prompt "A vertical comic about a banana detective" \
--pw "test-banana" \
--model-id "nano-banana-2" \
--aspect-ratio "1:8" \
--image-size "2K" \
--layout-rows 7 \
--layout-columns 1
Attach local images by file path:
node /absolute/path/to/banana_studio_generate.mjs \
--prompt "Use the references for style and character design" \
--pw "test-banana" \
--reference-image /abs/path/ref-1.png \
--reference-image /abs/path/ref-2.jpg \
--layout-guide /abs/path/layout-guide.png
Attach remote reference images by URL:
node /absolute/path/to/banana_studio_generate.mjs \
--prompt "Use the portrait as the character reference" \
--pw "test-banana" \
--reference-image "https://cdn.example.com/libai.png"
Storyboard layout is customizable; you are not limited to the bundled examples.
assets/jingyesi-4-panel.md and assets/jingyesi-1x4.md are only example templates.
You can create your own storyboard file with any reasonable layoutRows, layoutColumns, canvas size, panel content, and captions.
You can also override a storyboard file's default layout at runtime with --layout-rows and --layout-columns.
Output
The script:
streams backend status to stderr
saves the generated image under ./storage/skill-runs/... by default
writes a result.json summary beside the image
in storyboard mode, writes a captioned final result.png rendered by the tool instead of relying on the model to paint text
also writes a zip asset package beside the image, including result.png, result.json, request.payload.json, and extracted reference/layout images
prints a concise JSON summary to stdout
Notes
If --pw is omitted, the script reads BANANA_STUDIO_PW.
If BANANA_STUDIO_PW is also omitted, the script loads .env from the skill directory and reads BANANA_STUDIO_PW.
If --api-base-url is omitted, the script reads BANANA_STUDIO_API_BASE_URL and falls back to http://127.0.0.1:3001.
The local skill config uses standard dotenv / .env format with KEY=value entries such as BANANA_STUDIO_PW=....
The script defaults to modelId: "nano-banana-2" and frontend-like image defaults when you do not provide them.
--reference-image and --layout-guide accept either a local file path or a remote image URL.
In payload.json, referenceImages and layoutGuideImage may be passed as embedded frontend-style objects, local paths, or remote image URLs.
When the user asks for per-panel captions, prefer --storyboard-file so captions are rendered by the tool and the generated images themselves can stay text-free.
In storyboard mode, the script derives a near-matching cell aspect ratio from canvasWidth / layoutColumns and canvasHeight / layoutRows, then adds a prompt hint to keep the subject inside a safe area so final export cropping is reduced.
In storyboard mode, --layout-rows and --layout-columns override the file's layout and reflow cells by order, which avoids accidental 2x2/1x4 mismatches.
If storyboard coordinates fall outside the configured layout or collide, the script now throws an error instead of silently clamping them.
The printed result JSON includes assetPackagePath so callers can return a downloadable zip together with result.png and result.json.
When the user gives a reference image URL in natural language, treat it as a referenceImages input rather than leaving it only inside the prompt text.
When Codex uses this skill, resolve scripts/banana_studio_generate.mjs relative to the skill directory instead of assuming a repo-local skills/ folder.
