═══════════════════════════════════════════════════════════════════════════════ 🎨 GEMINI 3.1 FLASH IMAGE GENERATION ═══════════════════════════════════════════════════════════════════════════════

Supports: • Gemini 3.1 Flash Image (Nano Banana 2) - Fast, high-volume, 512px-4K

🌟 KEY CAPABILITIES: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ✓ High-Resolution Output: 512px, 1K, 2K, 4K ✓ Advanced Text Rendering: Legible text in infographics, diagrams, menus ✓ Reference Images: Up to 14 images (10 objects, 4 characters) ✓ Grounding: Google Web Search & Image Search ✓ Thinking Mode: Configurable reasoning (minimal or high) ✓ Transparent Backgrounds: one flag → ready-to-use alpha PNG/WebP cut-outs. See below — it just works. ✓ SynthID Watermarking: Invisible watermark on all images

🚀 WHY GEMINI 3.1 FLASH IS DIFFERENT: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ This isn't your old image generator. Gemini 3.1 Flash has LIVE ACCESS to Google Search and Image Search - it can find actual references for ANYTHING.

Examples: • "Way of Wade 12 latest colorway" → model finds the real shoe online • "Tony Hawk doing a kickflip" → model finds actual Tony Hawk photos • "iPhone 16 Pro Max" → generates the REAL device, not a guess • "Taylor Swift at the 2024 VMAs" → finds real reference images

Don't over-prompt! Simple descriptions work best. The model COOKS.

📋 PARAMETERS: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

► prompt (required, str): The text description. Be descriptive and specific. TIP: Less is more. "Tony Hawk kickflip" > "A man with long blonde hair wearing a skateboarding helmet doing a trick on a skateboard"

► enable_google_search (optional, bool, default: False): Enable Google Web Search for real-time data grounding. USE THIS FOR: Products, people, events, places, anything that exists NOW. The model will search for current info and generate ACCURATELY.

► enable_image_search (optional, bool, default: False): Enable Google Image Search for visual context. USE THIS FOR: Any visual reference - the model finds real images to work from. This is the "secret sauce" - it can reference actual photos of people, products, art, anything on the web.

► aspect_ratio (optional, str, default: "1:1"): OPTIONS: "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"

► image_size (optional, str, default: "2K"): OPTIONS: "512px", "1K", "2K", "4K" • "512px": Fastest, lowest cost (0.5K) • "2K": Recommended balance

► output_format: "png" (default), "jpeg", "webp"

► reference_image_paths (optional, str | list[str]): Path(s) to up to 14 reference images (10 objects + 4 characters). Accepts either a single path string (e.g. "/path/to/ref.png") or a list of path strings (e.g. ["/a.png", "/b.png"]).

► thinking_level (optional, str, default: "minimal"): Controls reasoning effort: "minimal" (fast) or "high" (best quality, slower). PRO TIP: Use "high" when using Google/Image search for best results.

🧠 THINKING MODE: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Gemini 3.1 Flash uses reasoning to refine composition before generating. Use thinking_level to balance quality vs latency: • minimal: Fastest, basic prompts • high: Best quality for complex prompts, slower PRO TIP: Use "high" thinking when using Google/Image search for best results.

🪟 TRANSPARENT BACKGROUNDS — JUST SET transparent_background=True: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ✅ THIS WORKS GREAT. Set transparent_background=True and you get back a ready-to-use transparent PNG/WebP with a real alpha channel — no extra tools, no manual masking, no follow-up steps. Use it directly.

Behind the scenes this uses a TWO-PASS DIFFERENCE MATTE: the subject is rendered once on a pure WHITE background, then that image is edited to a pure BLACK background, and the two frames are combined to recover a true (fractional) alpha channel. This costs a second model call (≈2x tokens/latency) but gives materially better edges than color-keying — clean soft edges, glow, glass, and shadows, with no green halo. You don't prompt for transparency; you just ask for it.

► transparent_background (bool, default: False): Flip to True to get the transparent cut-out. That's the whole API.

► alpha_output_format (str, default: "png"): Alpha output format: "png" (default) or "webp".

► preserve_original (bool, default: True): Also keeps the pass-1 (white-background) image next to the cut-out; set False for just the transparent file.

Each image returns transparent_path (your alpha file) plus background_removed, aligned, alignment_error and post_processing_warnings so you can confirm the cut succeeded. It nails crisp-edged subjects and soft glow/glass; the one failure mode is the edit pass drifting the subject (flagged via aligned=false) — regenerate if the edges look ghosted.

📤 RESPONSE FORMAT: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Returns JSON with: { "success": true, "images": [ { "path": "/path/to/image.png", "size": 12345 } ], "metadata": { "thinking_level": "minimal", "grounding_metadata": {...} } }

⚠️ IMPORTANT - AI ASSISTANT INSTRUCTIONS: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

1. Parse JSON to get file path: result["images"][0]["path"] (when transparent_background=True, use result["images"][0]["transparent_path"]).

2. Inform user of the EXACT file path.

3. Open image in native OS viewer using Bash:

   • macOS: open "/path/to/image.png"

   • Linux: xdg-open "/path/to/image.png"

   • Windows: start "" "/path/to/image.png"

💡 Need a transparent cut-out? Don't hand-mask or reach for another tool — just call this tool with transparent_background=True and use the returned transparent_path. It's built for exactly that.

Generate multiple images from a list of prompts efficiently.

Processes prompts in parallel batches for optimal performance. All images share the same generation settings.

═══════════════════════════════════════════════════════════════════════════════ 🍏 APP ICON & LOGO GENERATOR (square · transparent · ready for .iconset) ═══════════════════════════════════════════════════════════════════════════════

Use THIS tool — not generate_image — whenever the user asks for an app icon, application icon, .icns, .iconset, macOS/iOS/Android icon, favicon, logo, logomark, or brand mark. It is purpose-built for that job and removes every way to get it wrong.

🔒 WHAT IS FORCED (you cannot override these — by design): ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ • TRANSPARENT background — ALWAYS. Every result is a real alpha-channel PNG cut-out. There is no opaque-background option, because an icon or logo with a baked-in rectangle behind it is wrong. The transparent file path comes back as "transparent_path". • 1:1 SQUARE — ALWAYS. Every app icon is square; there is no aspect-ratio knob to get wrong. • 1K (1024px) — ALWAYS. This is the master size every .iconset slice and store listing is downscaled from. • PNG — ALWAYS. The lossless alpha format icons ship in. • CUT-OUT ONLY — ALWAYS. Only the transparent PNG is written; the pass-1 (white-background) original is never kept.

⛔ HOW TO WRITE THE PROMPT (READ THIS — the tool enforces it): ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ This tool IS the app-icon maker. Your prompt must describe ONLY the subject/artwork — NOTHING about the deliverable. Do NOT write "app icon", "application icon", "make an icon of", "logo of", "squircle", or anything about output/format/shape. The tool turns your subject INTO the icon.

WRONG: "an app icon of a blue magnifying glass over a network" RIGHT: "a glowing electric-blue magnifying glass over a network graph"

If your prompt contains "app icon", "logo", "favicon", "squircle" (or similar framing), the tool will REJECT the call and make you rewrite it. Just describe the picture.

The ONLY exception is when one of those words is genuinely PART OF THE SUBJECT you are depicting — e.g. a neon sign that literally reads "LOGO", or a picture OF a favicon. In that rare case, set allow_icon_words_in_prompt=True to bypass the check. Do NOT use it just to sneak deliverable-framing past the guard.

📋 PARAMETERS (what you DO control): ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ► prompt (required, str): Describe ONLY the subject/artwork itself — see the rule above. Don't ask for a background, a rectangle, a drop shadow, or a presentation surface. A bold, simple, single focal form reads best at small sizes.

► reference_image_paths (optional, str | list[str]): Brand/style reference image path(s), up to 14 (e.g. an existing logomark or palette to stay consistent with).

► enable_google_search / enable_image_search (optional, bool): Ground the design in real brand/product references found on the web.

► thinking_level (optional, str, default: "high"): "minimal" or "high". Defaults to "high" — icons reward the extra composition reasoning.

► allow_icon_words_in_prompt (optional, bool, default: False): Escape hatch for the prompt guard. Leave False. Set True ONLY when a word like "logo"/"favicon" is literally part of the subject you are depicting (e.g. a neon sign reading "LOGO"), not framing of the deliverable. Misusing this to bypass the guard defeats the point.

📤 RESULT / NEXT STEPS: ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Use result["images"][0]["transparent_path"] — that is the square, transparent 1024px PNG. Tell the user the exact path and open it in the native OS viewer (macOS: open "<path>"). To ship a macOS app, drop it into a .iconset directory and run iconutil -c icns <name>.iconset. For an iOS App Store upload, flatten onto an opaque background first (Apple rejects icons that contain an alpha channel).