Nano Banana Pro

What it is

Nano Banana Pro is Google's Gemini 3 Pro Image model: 4K native output, multi-image reference input, "thinking mode", genuinely good multi-language text rendering. It is also closed, so it will never run on your card - this node is how it gets into your graph, calling KIE's marketplace API instead of Google directly.

That's a deliberate choice, and the motive is money. When ComfyUI added native Nano Banana Pro support, people looked at the per-image pricing and went and built their own node packs against a reseller instead. Community comparisons through KIE put Pro at roughly 18 credits (~$0.09) for 1K/2K and 24 (~$0.12) for 4K, against roughly $0.15 and $0.30 on the first-party Fal-backed route - about 40% cheaper at 1K/2K and 60% at 4K. If you generate 2×2 or 3×3 grids, that's not a rounding error.

What you're giving up is the officially vetted path and, for some, the comfort of first-party billing. Everything in docs/knowledge/external-api-nodes.md about API nodes applies: your inputs leave the machine, a filter you can't remove sits inside the model, and cost is metered per call.

How it works

The node is generated from KIE's documented request schema for the Pro endpoint: four widgets, all optional apart from the prompt. Connect an IMAGE and the pack uploads it to KIE's storage and sends the resulting URL - there is no URL field to fill in. The task is submitted, polled until it finishes or timeout_seconds runs out, and the finished image is downloaded and returned as a ComfyUI IMAGE.

Inputs and outputs

Required:

• prompt (STRING, max 10,000 characters) - describe the image you want. The author's shipped default is a six-panel comic-poster demo; clear it before your first run unless you want to pay for a banana wearing shades.

Optional:

• image_input (IMAGE) - the reference image. Pro's reputation is built on handling several references coherently; a batch here is your multi-reference path.
• resolution - 1K, 2K, 4K. 4K is the headline capability and the expensive one. Worth it for a final asset, wasteful for a draft.
• aspect_ratio - eleven options, 1:1 default, plus auto. Set it deliberately when the output feeds a layout.
• output_format - png or jpg.
• timeout_seconds (1200 default) and callback_url (leave empty).

Outputs: image, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left.

Install

Manager: search KIE-Nodes-Next by felipederosilva - several KIE packs live in that search result, so check the repo - then install and restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next

Open Settings → KIE.ai Nodes Next → Connection and paste your KIE.ai API key. The pack validates it before replacing a saved key and reports a masked fingerprint, a ✅ CONNECTED status and your balance. KIE_API_KEY environment variable overrides everything for headless installs. Requirements are requests, Pillow, numpy, PyYAML - no torch surgery.

Where people get burned

4K is the price cliff. It's the reason to use Pro and the reason your credit balance empties. Draft at 1K; promote only what survives review. If you can get a good 1K from a cheaper model and just need more pixels, a local upscaler is often the smarter spend (see docs/knowledge/upscaling.md).

"Powerful, unimaginative" is the honest critique. That's how this model line gets characterised in community discussion, alongside "massive hidden biasing" - it's excellent at executing a detailed brief and mediocre at being interesting on its own. Bring a plan; use a local model for wild ideas and this one to land them.

Filters you can't patch. Pro's refusals come from Google. Nothing in the node changes them.

Model churn on the KIE side. Endpoints appear and retire; if a run starts failing with no workflow change, run KIE • Refresh Model Catalog and check the model is still listed.

The demo prompt again. Same trap, same fix: clear it.

Read credits_consumed like a receipt. Every successful run prints credits spent and the live balance on the node. That's not decoration - it's the only cost telemetry you're going to get.