KIE Nano Banana Pro (Image)

Nano Banana Pro is Google's flagship Gemini 3 Pro image model - 4K native output, up to 8 reference images, and a "thinking" pass over your prompt before it generates. It's also, as of mid-2026, one of the best closed image models you can't run locally. This node is the reason a lot of ComfyUI users reach for the whole gateway/ComfyUI-Kie-API pack: it's the origin story.

The pack's author built it after realizing that calling Nano Banana Pro through the Kie.ai marketplace was dramatically cheaper than the native ComfyUI/Fal integration. At the time of the launch post: roughly $0.09–0.12 per image through Kie versus $0.15–0.30 through the Fal-backed native node - around 40–60% cheaper, most of it at 4K. If you generate grids (2×2 or 3×3 at 4K is the author's own workflow), that difference stops being trivia and starts being a budget.

How it works

Same architecture as the rest of the pack: upload prompt + reference images to the Kie.ai job API, poll until the image renders, decode into a ComfyUI IMAGE tensor. The model's "thinking mode" happens server-side before generation, which is exactly why a one-line prompt can come back with a composition that actually respects your references.

Inputs that matter

• prompt - the only required input. Multiline, plain descriptive language works best.
• images - optional reference images, up to 8 for Pro (vs 14 on Nano Banana 2). For identity work - "keep this face" - wire one in.
• aspect_ratio - auto by default.
• resolution - 1K, 2K, or 4K. Pro is 4K-native; that's where it shines and where the credits go.
• output_format - png by default (Pro defaults to lossless, unlike Nano Banana 2's jpg).
• log - console logging, on by default.

Output: a single image (IMAGE) that plugs into SaveImage, GridSlice, or anything else.

The workflow it was built for

The pack's example "Banana Pro Grid" workflow is the canonical use: an LLM analyzes an input (face reference + source image) and produces a set of prompt variations, each passed to a Nano Banana Pro run, generating a 2×2 or 3×3 grid in one call. A GridSlice node then cuts the grid into individual images, and each becomes a first frame for downstream video. That's the loop - iterate cheaply on grids, slice, animate. If that's your pipeline, this node is the engine.

Install and key setup

Manager (search "ComfyUI-Kie-API") or:

cd ComfyUI/custom_nodes
git clone https://github.com/gateway/ComfyUI-Kie-API

Restart, then config/kie_key.txt (copy from config/kie_key.example.txt) with your Kie.ai key. No models to download, no GPU requirements - the only real dependency is a funded Kie account.

Where people get burned

• 4K costs 4K. The pricing jump at 4K is the one place the cheap-vs-native gap closes. For iteration, run 1K/2K; save 4K for finals.
• It's still a Google safety filter. The KB's notes on Nano Banana apply: aggressive IMAGE_SAFETY filtering and firm IP restrictions. This is a hosted flagship, not a local LoRA playground.
• The thinking pass is slower. Pro "reasons" before it renders, so calls take longer than Nano Banana 2. With the 2000s timeout you're fine, but don't mistake a 30-second render for a hang.
• Check your balance early. The pack's "Get Remaining Credits" node exists because API credit surprises are the #1 support question for this kind of node. Verify the key, then generate.

For reference-heavy, quality-first image work inside ComfyUI, this is the node the pack is named after, and it's still the one to beat.