arkennemasis fal · Nano Banana Pro · $0.15-0.30/image

Nano Banana Pro is Gemini 3 Pro Image: 4K native output, thinking mode as standard, and the best in-image text rendering in the closed lineup. $0.15 an image, doubling to $0.30 at 4K. It's the expensive one in this family and, for a specific job, the one that's actually worth it.

When to pay the 15 cents

Text inside the image. That's the honest answer. Local models have improved at lettering but they still can't be trusted with a headline, a label, a menu, a poster, or a legible UI mockup - and this model can. If your shot has words in it that a human will read, paying 4× the base model's price is not a waste; it's the difference between a concept and a deliverable.

The other case is resolution. 1K is the default here, and it scales to 2K and 4K natively rather than by upscaling - so if the output is going to be printed, cropped hard, or used as a plate for a video model, the extra $0.15 buys real pixels. Think about it this way: 4K output at 30 cents is still cheaper than a 4K upscale pass on a rented GPU, and it doesn't invent detail that wasn't there.

The settings

• prompt - the scene. Long, specific prompts work well here; Pro reasons about the composition before drawing.
• resolution - 1K (1×), 2K (2×)... and 4K at 2× per fal's published multiplier, i.e. $0.30. The node's live badge multiplies as you change it, so watch that number rather than doing arithmetic in your head.
• aspect_ratio - defaults to 1:1 on this node, and unlike the Nano Banana 2 family it doesn't advertise the extreme 4:1/8:1 ratios. Set it deliberately.
• num_images - 1 to 4, each billed at the multiplied rate.
• enable_web_search - off by default, +$0.015 when on. For anything referencing current products, this is cheap.
• system_prompt - a standing instruction that steers persona and style across the request. Good for "no text, no watermarks" or a house style.
• safety_tolerance - 1 (strict) to 6 (loose), default 4. Google's IMAGE_SAFETY filtering applies and the January 2026 policy tightened famous-IP and celebrity generation. No loosening trick exists at this layer; the filter is in the model.
• seed - an integer, -1 for the model's choice.
• limit_generations - on by default, caps each prompting round at one image and ignores any number-of-images instruction in your prompt.

Note there's no thinking_level widget here - Pro is the thinking model, so it's simply on.

Outputs

images - batchable IMAGE output, ready for an upscaler, a video model's first frame, or a save node. image_paths - the files, under output/fal/fal-ai_nano-banana-pro/. description - the model's own text about what it produced. info - JSON with the request id, cost estimate, saved files and fal's raw answer.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Restart, or use ComfyUI-Manager (search arkennemasis). The pack's requirements.txt is one line - httpx, for its Codex nodes - and the fal provider uses the standard library, so no dependency surprises. Then:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

in the .env next to run_nvidia_gpu.bat. There's no key box on the node, and that's a feature: a typed key would be saved into every workflow JSON you share.

Where people get burned

• You spent 30 cents experimenting. Draft on the cheap node ($0.0398) or Nano Banana 2 ($0.08), then move the prompt you liked to Pro and spend the real money once.
• Four 4K images cost $1.20. Multipliers and num_images compound. The badge shows it; the receipt in info confirms it.
• Square images again. aspect_ratio defaults to 1:1 here, not auto.
• One image instead of four. limit_generations is on by default.
• A blocked prompt with nowhere to go. Google refuses, and this family's filters are the strictest in the line. That's the standing argument for keeping a local model on the canvas for anything the hosted filter won't do.