nano-banana-image-gen

name	nano-banana-image-gen
description	Use when the user asks to generate or edit images with Nano Banana / Nano Banana Pro (users often call these models 'banana'). Nano Banana (aka Nano Banana 2) maps to `gemini-3.1-flash-image-preview`; Nano Banana Pro maps to `gemini-3-pro-image`.

Nano Banana Image Generation Skill

Generate or edit images using Google's Gemini native image generation. Naming note: Nano Banana (aka Nano Banana 2) maps to gemini-3.1-flash-image-preview, and Nano Banana Pro maps to gemini-3-pro-image (currently exposed in this CLI as gemini-3-pro-image-preview).

When to use

• Generate a new image (concept art, product shot, cover, hero, sticker, icon)
• Edit an existing image (style change, object removal, background replacement, text localization)
• Multi-image composition (combine up to 14 reference images)
• Search-grounded generation (real-time data like weather, events, charts)

Model selection

• gemini-3.1-flash-image-preview (Nano Banana / Nano Banana 2, default): best speed/quality balance, supports Thinking + Search Grounding, and supports 512px low-latency mode. Best with <=3 input images.
• gemini-3-pro-image-preview (Nano Banana Pro / gemini-3-pro-image): professional asset production, advanced text rendering, Thinking mode, up to 4K, Google Search grounding. Supports up to 14 input images (5 humans high-fidelity, 6 objects high-fidelity).

Use Pro when the user needs: highest text rendering quality, complex multi-image composition, or explicitly requests "pro" / "high quality".

Decision tree

• If the user provides input image(s) or says "edit/retouch/change/modify/localize" -> edit
• If the user needs Search-grounded imagery (current weather, live data) -> generate with --google-search
• Else -> generate

Workflow

1. Decide intent: generate vs edit (see decision tree).
2. Collect inputs: prompt, input images (if any), constraints, text to render (verbatim).
3. Choose model: Flash (default) or Pro (see model selection above).
4. Augment prompt into a structured spec (see prompt augmentation below); do not invent new creative requirements.
5. Run the bundled CLI (scripts/gemini_image_gen.py). See --help for commands and flags.
6. Inspect outputs and validate: subject, style, composition, text accuracy.
7. Iterate: make a single targeted change, re-run, re-check.
8. Save final outputs and note the final prompt + flags used.

Output conventions

• Write final artifacts under output/imagegen/ when working in this repo.
• Use --out to control output path; keep filenames stable and descriptive.

Dependencies

cd {SKILL_BASE_DIR} && uv run --with google-genai --with google-auth --with pillow scripts/gemini_image_gen.py generate --prompt "test" --dry-run

Defaults & rules

• Model: gemini-3.1-flash-image-preview unless user asks for Pro or needs Pro features.
• Aspect ratio: 1:1 (default). Supported: 1:1, 1:4, 1:8, 2:3, 3:2, 3:4, 4:1, 4:3, 4:5, 5:4, 8:1, 9:16, 16:9, 21:9 (1:4/1:8/4:1/8:1 are Nano Banana 2 only).
• Resolution: 1K (default). Supported: 512px, 1K, 2K, 4K; 512px is Nano Banana 2 only.
• Response modalities: TEXT,IMAGE by default. Use --image-only for image-only output.
• Output format: PNG (default).
• Prefer the bundled CLI over writing new scripts.
• Never modify scripts/gemini_image_gen.py. Ask the user if something is missing.

Prompt augmentation

Reformat user prompts into a structured spec. Only make implicit details explicit; do not invent new requirements.

Template (include only relevant lines):

Use case: <taxonomy slug>
Primary request: <user's main prompt>
Scene/background: <environment>
Subject: <main subject>
Style/medium: <photo/illustration/3D/etc>
Composition/framing: <wide/close/top-down; placement>
Lighting/mood: <lighting + mood>
Color palette: <palette notes>
Text (verbatim): "<exact text>"
Constraints: <must keep/must avoid>
Avoid: <negative constraints>

Use-case taxonomy

Generate:

• photorealistic-scene -- candid/editorial photos with real texture and natural lighting
• product-mockup -- product/packaging shots, catalog imagery
• ui-mockup -- app/web interface mockups
• infographic-diagram -- diagrams/infographics with structured layout and text
• logo-brand -- logo/mark exploration, minimal flat design
• illustration-sticker -- icons, stickers, kawaii art, children's book art
• stylized-concept -- style-driven concept art, 3D renders
• minimalist-negative-space -- backgrounds for websites/presentations with text overlay space
• search-grounded -- real-time data visualization (weather, charts, events)

Edit:

• style-transfer -- apply reference style to new subject
• object-edit -- remove/replace specific elements
• text-localization -- translate/replace in-image text
• multi-image-composition -- combine multiple reference images
• sketch-to-render -- drawing/line art to photoreal render

Prompting best practices

• Describe the scene narratively; do not just list keywords.
• Use photography language for photorealism (lens, lighting, framing).
• For text in images: quote exact text, specify font style descriptively, use Pro model.
• Nano Banana 2 also supports Thinking and Search Grounding; use it when speed matters.
• For Pro model: let Thinking work; complex prompts get better results.
• For Search grounding: describe what real-time data to visualize.
• For edits, repeat invariants ("change only X; keep Y unchanged").
• Iterate with single-change follow-ups.

Reference map

• references/cli.md: how to run generation/edits via scripts/gemini_image_gen.py (commands, flags, recipes).
• references/gemini-api.md: API-level parameters, models, authentication modes, constraints.
• references/prompting.md: prompting principles, strategies, and templates from Google's official guide.
• references/sample-prompts.md: copy/paste prompt recipes for each use-case taxonomy slug.
• references/vertex-setup.md: how to configure Vertex AI credentials (service account, env vars, IAM roles).