Nano Banana — Gemini Image Generation

# Nano Banana — Gemini Image Generation CLI Generate images using the `nano-banana` CLI powered by Gemini 3.1 Flash (default) and Gemini 3 Pro. Free with a paid Google account. ## Gemini CLI vs fal.ai Pipeline — When to Use Which | Need | Use This Skill (nano-banana CLI) | Use ad-creative Skill (fal.ai) | |------|----------------------------------|-------------------------------| | Quick standalone image | Yes | No | | Transparent asset / icon | Yes (`-t` flag) | No | | Style transfer / image editing | Yes (`-r` flag) | No | | Mockup / UI screenshot | Yes | No | | Carousel slide illustration | Yes | No | | **Jay's face in the image** | **No** — use ad-creative | **Yes** (Flux LoRA) | | **Full 5-step ad pipeline** | **No** — use ad-creative | **Yes** (Scene → LoRA → Composite → Text → QA) | | **Smart Swap / Body Swap** | **No** — use ad-creative | **Yes** | **Rule of thumb:** If it needs Jay or the ad pipeline → `ad-creative`. Everything else → `nano-banana`. ## Quick Reference ```bash # Basic generation (1K, Flash model) nano-banana "your prompt" # With options nano-banana "prompt" -o filename -s 2K -a 16:9 -d output/dir # Transparent asset nano-banana "robot mascot" -t -o mascot # Style transfer / edit nano-banana "make background white" -r input.png -o result # Pro model (highest quality) nano-banana "detailed hero image" --model pro -s 2K ``` ## Core Options | Option | Default | Description | |--------|---------|-------------| | `-o, --output` | `nano-gen-{timestamp}` | Output filename (no extension) | | `-s, --size` | `1K` | `512`, `1K`, `2K`, or `4K` | | `-a, --aspect` | model default | `1:1`, `16:9`, `9:16`, `4:3`, `3:4`, `4:5`, `5:4`, `21:9` | | `-m, --model` | `flash` | `flash`/`nb2`, `pro`/`nb-pro`, or any model ID | | `-d, --dir` | current directory | Output directory | | `-r, --ref` | — | Reference image(s), can use multiple | | `-t, --transparent` | — | Green screen → background removal (needs ffmpeg + imagemagick) | | `--costs` | — | Show cost tracking summary | ## Model Selection Guide | Model | Alias | Best For | Cost (1K) | |-------|-------|----------|-----------| | Gemini 3.1 Flash | `flash`, `nb2` | Default. Speed, batch generation, concepts, mockups | ~$0.067 | | Gemini 3 Pro | `pro`, `nb-pro` | Hero images, style transfer, thumbnails, final assets | ~$0.134 | **Free with paid Google account** — use Flash for iteration, Pro for finals. ## Lead Gen Jay Presets ### Brand-Aligned Prompts **Dark mode SaaS dashboard:** ```bash nano-banana "premium SaaS dashboard, dark mode with near-black #0D0D0D background, hot pink #ED0D51 accent highlights on charts and buttons, clean minimal UI, analytics view" -a 16:9 -s 2K -o dashboard ``` **Marketing hero image:** ```bash nano-banana "cinematic tech workspace, dramatic lighting, laptop showing automation dashboard, professional atmosphere, shallow depth of field" --model pro -s 2K -o hero ``` **Transparent carousel icon:** ```bash nano-banana "flat illustration of a robot arm automating emails, clean vector style, bold colors" -t -s 1K -o carousel-icon ``` **Product mockup:** ```bash nano-banana "MacBook Pro on clean desk showing SaaS dashboard with analytics, minimal setup, soft lighting, editorial product photography" --model pro -s 2K -a 16:9 -o product-mockup ``` ### Ad Creative Sizes | Format | Flag | Use Case | |--------|------|----------| | 1:1 (1024x1024) | `-a 1:1` | Feed ads, social posts | | 4:5 (1080x1350) | `-a 4:5` | Instagram feed (max real estate) | | 9:16 (1080x1920) | `-a 9:16` | Stories, Reels, TikTok | | 16:9 (1920x1080) | `-a 16:9` | YouTube thumbnails, hero images | ## Output Directory Conventions Route output to project structure using the `-d` flag: ```bash # Ad scene backgrounds (before pipeline Steps 4-5) nano-banana "..." -d output/ads/ # Carousel slide assets nano-banana "..." -t -d output/carousel-assets/ # YouTube thumbnails nano-banana "..." -a 16:9 -d output/thumbnails/ # General photos / mockups nano-banana "..." -d output/photos/ ``` ## Pipeline Integration Nano-banana output can feed into the ad-creative pipeline at Steps 4-5: 1. **Generate scene** with nano-banana → use as base image 2. **Apply text overlay** with `addAdHeadline()` (Sharp + SVG from `scripts/generate-cold-email-ads.mjs`) 3. **QA with Claude Vision** via `analyzeWithVision()` for scoring **Transparent assets** feed directly into carousel `NotebookBackground` components as slide visuals. See `references/integration-guide.md` for detailed integration patterns. ## Key Workflows ### Transparent Assets (Icons, Sprites, Logos) ```bash nano-banana "robot mascot character" -t -o mascot nano-banana "minimalist tech logo, bold geometric" -t -o logo nano-banana "pixel art treasure chest" -t -s 512 -o chest ``` ### Reference Image Editing ```bash # Edit existing image nano-banana "change background to pure white" -r dark-ui.png -o light-ui # Style transfer from multiple references nano-banana "combine these two UI styles" -r style1.png -r style2.png -o combined ``` ### Batch Concept Generation ```bash # Quick low-res concepts first nano-banana "automation dashboard concept A" -s 512 -o concept-a nano-banana "automation dashboard concept B" -s 512 -o concept-b # Then upscale the winner nano-banana "refined version of concept A with more detail" -r concept-a.png --model pro -s 2K -o final ``` ## Image Size Limits (Context Safety) - All downloaded/scraped reference images used with `-r` must be max 1000px — resize before use: `sips --resampleHeightWidthMax 1000 <file>` - NEVER use Read tool to view images — use `open` for Preview.app or `sips -g pixelWidth -g pixelHeight` for dimensions - nano-banana output at 1K (1024px) is already safe — only external/scraped images need resizing - This prevents Claude Code context crashes from the 2000px many-image limit ## Cost Tracking ```bash nano-banana --costs # Shows total spend, per-model breakdown ``` Logs stored at `~/.nano-banana/costs.json`. ## Related Skills | Skill | When to Use | |-------|-------------| | `ad-creative` | Jay's face, full 5-step ad pipeline, Smart Swap | | `frontend-design` | UI components that use generated assets |