Nano Banana Pro (Text to Image) 🍌

This is the other half of the ComfyUI_NBP_Unleashed pack: text-to-image with Google's Nano Banana Pro through fal.ai's paid API. No image inputs, no local model, no VRAM - just a prompt, a bill, and a result. If you're here because the edit node caught your eye, note this one up front: it has no safety_tolerance knob. The censorship question that draws people to this pack only has an answer on the edit node, and even there it's partial. Text-to-image gets whatever default filtering fal applies, full stop.

Why you'd reach for it

Nano Banana Pro is Google's flagship closed image model - native 4K output, strong prompt following, and unusually clean text rendering. The community's honest take is that it's technically impressive but heavily censored and "unimaginative" at the edges; it wins on competence, not creativity. Inside ComfyUI, this node is for the workflows that genuinely need that closed-model quality: generating 4K grids or concept boards to slice up for downstream video work, producing reference sheets, or batching ideas when you want the best available prompt adherence per dollar. If your job can be done by a local model you already have, it's free - reach for this when you need the quality gap or the 4K resolution and are willing to pay per image for it.

How it works

Mechanically it's simple and dependable. The node takes your prompt, builds a payload, POSTs to fal's synchronous fal-ai/nano-banana-pro endpoint, polls until the result lands (default timeout 300 seconds), then downloads the returned images into normal ComfyUI IMAGE tensors. Your API key comes from the FAL_KEY environment variable, or you can drop it directly into the node's fal_api_key field and skip the env var. The whole call happens inside the node, so it slots into a graph like any local generator - you just need to remember every run is a real-money cloud request.

The inputs that matter

Only one is required:

• prompt - multiline, and be specific. This model follows long, detailed instructions well, which is part of what you're paying for.

The rest you can mostly leave alone, but a few are worth knowing:

• num_images - 1 to 4, default 1. The API can return multiple per call.
• limit_generations - force exactly one image per round, even if num_images is higher. Handy for keeping costs predictable.
• resolution - 1K / 2K / 4K. 4K costs twice as much, so don't leave it on 4K by accident.
• aspect_ratio - 11 options including auto and the usual suspects (1:1, 16:9, 9:16, 4:3, 21:9…). Default is 1:1.
• seed - -1 (default) means random; set a positive value to reproduce a result.
• enable_web_search - lets the model search the web before generating. Costs an extra $0.015 per call, and usually you don't need it.

There's also output_format (png/jpeg/webp) and timeout (30–600s).

Outputs

Two: images - an IMAGE list, so wire it into SaveImage or further processing - and description, a STRING of the model's caption of what it made. When a prompt gets filtered, the description sometimes carries the refusal reason, so it doubles as a diagnostic.

Installing it

Same as the edit node - it's the same pack, so this is a shared setup.

cd ComfyUI/custom_nodes
git clone https://github.com/digital-divas-admin/ComfyUI_NBP_Unleashed
pip install requests

Set your key before starting ComfyUI:

export FAL_KEY="your-fal-api-key-here"

Restart and look under fal.ai/Nano Banana Pro. ComfyUI Manager finds it if you search "ComfyUI_NBP_Unleashed". The only dependency is requests - PIL, torch and numpy already ship with ComfyUI - and there are no model files to download. Get a key at https://fal.ai/dashboard/keys.

Common issues

Filtered prompts are the main gotcha, and since there's no safety_tolerance on this node, you can't tune your way around them - a refusal comes back as an empty images array or an error, and the description output is your best clue why. The other trap is the bill: $0.15 per image at 1K/2K, $0.30 at 4K. If you're iterating hard, watch your resolution and count your runs; heavy users regularly shop around to resellers like Kie.ai for cheaper per-image rates on the same model. And if generation ever seems to hang, the 300s default timeout is usually the culprit - bump it toward 600 for long, complex prompts rather than assuming the node is broken.