KIE Nano Banana Pro (Image) [SL]

This is the node that started the whole pack. Nano Banana Pro is Google's flagship closed image model - 4K native, "thinking mode" before it draws, the best in-graph text rendering you can get without open weights - and there are no weights to download. KIE resells access to it, and this node is the bridge from your ComfyUI canvas to that billable endpoint.

Why use a reseller instead of ComfyUI's own official Nano Banana API node? Price. The pack's author documented the gap on Reddit: at 1K/2K, KIE ran roughly $0.09/image against the native node's ~$0.15, and at 4K it was ~$0.12 vs ~$0.30 - around 40-60% cheaper. If you batch, that difference is a real budget item, and it's the honest reason this whole ecosystem of third-party KIE packs exists. The tradeoff is you're one layer further from Google: your prompts and images go through a reseller's pipeline, so treat it accordingly.

The inputs that matter

• prompt - up to 10,000 characters. This is a Gemini-model, so it rewards natural-language instruction, not tag soup.
• images - optional reference images, up to 8. Leave it unwired and you get plain text-to-image; wire it up and it becomes a reimagining/edit node.
• aspect_ratio - wide enum from 1:1 to 21:9 plus auto (default), which lets the model pick from the content.
• resolution - 1K, 2K, or 4K. 4K is where the cost jumps.
• output_format - png or jpg.

Output is a single image tensor (BHWC float32 0–1), ready for SaveImage or anything downstream.

How it works

Upload your reference images as PNGs, create an async job for the nano-banana-pro model, poll until success, download, decode. The same KIE job machinery every node here uses, with built-in retry on transient 429/5xx errors. If you feed more than 8 images, the extras are silently dropped - keep batches at 8 or under.

Install

ComfyUI Manager search "Kie-API-Serverless", or:

cd ComfyUI/custom_nodes
git clone https://github.com/v3devv/ComfyUI-Kie-API-Serverless

Restart. No models, no extra dependencies - just your KIE key in the node's api_key field (this is the serverless fork, so the key lives in the node input rather than a config file).

Where people get burned

Two things, both predictable. First, cost creep: 4K per-call pricing plus your habit of re-running until a grid is perfect is how a KIE balance evaporates. The classic move - this is literally the workflow the original pack author demoed - is to generate a 2x2 or 3x3 grid in one call and slice it with the pack's GridSlice node instead of buying nine separate images. Second, the reseller layer means KIE's availability and pricing can drift; if a task fails, the failCode/failMsg in the error is what you take to support. And remember the moderation axis: this is Google's model under the hood, so anything Google refuses to draw, this node refuses - no local abliteration exists for a model you never owned.