Nano Banana Pro (fal)

"Nano Banana" started as an anonymous codename on a leaderboard, Google execs leaned into the joke with banana emojis, and now it's the official name of Gemini's native image generation. Nano Banana Pro is the flagship - Gemini 3 Pro Image - and it's the one you reach for when you need two things most local models are weak at: crisp, accurate text inside the image, and reliable multi-image compositing. This node runs it from ComfyUI, and unlike most nodes in this pack it hands you back a real IMAGE, so it drops straight into a normal graph.

What it actually does

Nothing runs locally. Your prompt (and any reference images) go to fal.ai, fal calls Google's model, and you get a picture back with no VRAM and no downloads - billed per image by fal. Nano Banana Pro's standout traits are 4K native output, strong prompt reasoning (it "thinks" before it renders), and multi-language text that actually comes out legible. The catch, and it's a real one, is censorship: Google's filters are aggressive and opaque, and every image carries an invisible SynthID watermark. If you value being able to make anything, this isn't your model. If you value clean typography and compositing, it's hard to beat.

The inputs that matter

• prompt (required): plain natural language. It rewards detail and it's unusually good at following layout and text instructions.
• images (optional): reference image(s) to edit or compose from. This is what turns it from pure text-to-image into an editor/compositor - feed it a product shot and describe the scene you want it dropped into.
• resolution: 1K, 2K, or 4K. 4K is the reason people use this model; it's also the priciest.
• aspect_ratio: eleven presets from 21:9 down to 9:16, plus auto.
• num_images: 1 to 4 per run - each one is billed, so it's a paid way to get options.

output_format (png/jpeg/webp) and sync_mode are there if you need them; defaults are fine. The single output is a native IMAGE - wire it into Save Image, Preview Image, an upscaler, or anything downstream that takes an image.

Installing it

Easiest: ComfyUI Manager → search ComfyUI-fal-API → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
cd ComfyUI-fal-API
pip install -r requirements.txt

Restart ComfyUI. There's no model to place in models/. The dependencies are fal-client and a fal API key: copy config.ini.example to config.ini in the pack folder and set the key under [API], or export FAL_KEY before launching. No key, no images.

Where people get burned

The most common surprise is a content-policy refusal - Google's filter is stricter and less predictable than most, and it's especially touchy about real people and famous IP. The pack surfaces fal's actual error text, so read it before assuming the node is broken; a refusal is the model declining, not a bug. Second, don't expect an uncensored, watermark-free file: the SynthID mark is baked in.

On cost, 4K × num_images of 4 adds up fast, so test at 1K and scale up once the prompt is right - identical calls are cached free, but any change re-bills. And on ComfyUI Windows Portable, an import error is fixed with .\python_embeded\python.exe -m pip install fal-client.