Nano Banana Pro (Gemini 3 Pro Image)

Nano Banana Pro (Gemini 3 Pro Image) is the headline node in Ametist3d's pack, and it does exactly what it says: pipes Google's flagship image model straight into your ComfyUI graph. One 4096×4096 4K render, no checkpoint, no VRAM, no controlnet pile - the diffusion happens on Google's servers and the result lands as a normal IMAGE tensor you can save, upscale, or feed into a LoRA training pipeline.

If you've seen people generating character sheets and LoRA datasets with Nano Banana, this is how they run it from inside ComfyUI. It's an API node: nothing is downloaded, and every generation costs money and needs a Gemini API key. It also speaks three models - gemini-3-pro-image-preview (Nano Banana Pro, the default), gemini-3.1-flash-image-preview (Nano Banana 2), and gemini-2.5-flash-image (Nano Banana) - and sanitizes settings per model so you can't send something the API will reject.

How it works

No local sampling here. The node builds a request with pydantic models and POSTs it to Gemini's generateContent endpoint over aiohttp: your prompt, base64-encoded input images (up to 14), optional mask, imageConfig for aspect ratio and resolution, a seed, and optional Google Search grounding tools. It then extracts the returned images back into tensors and - this is the nice part - enforces your chosen resolution locally, resizing the output to match the pixel budget even when the API returns input dimensions on an edit. Multi-turn refinement is handled by chat_metadata: wire the output back into the input and you keep editing the same image conversationally.

Inputs and outputs that matter

• prompt - the whole game. For edits, this is where you describe the change.
• model - 3 Pro Image by default; 3.1 Flash unlocks 0.5K, extended aspect ratios, thinking level, and image-search grounding.
• aspect_ratio - auto matches your input image or falls back to 1:1. Unsupported ratios are sanitized per model.
• resolution - pixel-area budget, 0.5K through 4K. Gotcha: 0.5K only works on 3.1 Flash (Pro silently bumps it to 1K), and 2K/4K only on Pro.
• seed - set it for reproducibility; 0 is random. Google treats the seed as "best effort," not a guarantee.
• images and mask - wire in a reference image for edit/refine, or a mask for inpainting (white areas get edited).
• response_modalities - IMAGE+TEXT (default) returns the image and a text caption/explanation; IMAGE skips the prose.
• system_prompt - comes prefilled with a "you must ALWAYS produce an image" instruction. Leave it alone unless you know why you're touching it.

Outputs: images (into Preview Image / Save / anything downstream), text (the model's written response), and chat_metadata (loop back into the input to keep refining).

Installing it

Shared with the pack's LLM node: Manager (search "Gemini3ProImagePreview") or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Ametist3d/ComfyUI_Gemini3ProImagePreview
cd ComfyUI_Gemini3ProImagePreview
pip install -r requirements.txt

Restart after. Real runtime deps are just aiohttp plus Pillow/numpy/torch - the listed google-generativeai is never imported.

The API key gotcha

The README says setx GEMINI_API_KEY - ignore it, the code never reads the OS environment. It loads the key from ~/comfyui_google_api_key.env first, then a .env file in the node folder containing GEMINI_API_KEY=.... So:

echo 'GEMINI_API_KEY=AIza...' > ~/comfyui_google_api_key.env

Restart, and you're live. A missing key raises a clear ValueError naming both paths.

What will bite you

Pricing is the first wall: Nano Banana Pro runs roughly $0.039–$0.24 per image depending on resolution, and Google Search grounding adds more. Second is censorship - the knowledge base notes the January 2026 safety-policy tightening, and this API is aggressively filtered; content it flags comes back empty. Relatedly, if you get a solid-black output tensor, that's the node's placeholder for "API returned no images" - usually a safety block, not a bug. And expectations: reviewers call Pro powerful but "unimaginative" and prone to hidden bias, so treat it as a tool, not a replacement for your local stack.