Kie Nano Banana 2

Nano Banana 2 is Google's image model - the speed-quality hybrid that dropped in February 2026 on Gemini 3.1 Flash. It's also completely closed. No weights to download, no GGUF to fit on your card; you only ever reach it through an API or ImageFX. KieNanoBanana2 is the ComfyUI front door to that API, through a third-party reseller called Kie AI. You hand it a prompt and an API key, and it hands you back a normal ComfyUI IMAGE tensor. Same graph, different planet.

The name isn't a lie, but it's worth being precise about: the model is Google's, the API is Kie's, and nothing runs on your machine. This is one of the API category nodes - the ones that let you lean on a closed-source model for the bits where open weights still fall short, then keep the result inside your local workflow. For Nano Banana specifically, that's a real pitch: the community's honest verdict is that it's powerful but "unimaginative" and aggressively filtered. If you want prompt adherence, clean text rendering, or reference-image fidelity without babysitting a local model, an API node like this is the pragmatic path. It's also handy when you're on a low-VRAM box and your "GPU" is somebody else's data center.

How it works

The node is a thin async client for Kie's API. When you run it, it does four things in order: uploads any reference images to Kie's base64 upload endpoint, creates a generation task (model: "nano-banana-2"), polls the task-detail endpoint until it succeeds, fails, or times out, then downloads the result images and decodes them back into a tensor. Polling starts every 2 seconds, backs off to 5, then 15 as a task drags on - sensible, since these jobs typically take well under a minute but can occasionally hang around.

One implementation detail worth knowing: reference images are converted to base64 PNGs and uploaded per image, in batch order. So if you feed it a batch of four reference images from elsewhere in your graph, all four get sent up and the model sees them in order - the mechanism for a simple multi-reference setup.

The inputs that matter

You only really touch a handful of these:

• prompt - the only required text. Leave it empty and the node raises immediately, so there's no silent no-op.
• api_key - the one gotcha. You can paste it in, but the README warns it then lives inside your workflow JSON, which is a leak if you share workflows. Better: set the KIE_API_KEY environment variable, or copy config.example.json to config.json in the pack folder and put the key there. Resolution order is node input → env var → config file.
• aspect_ratio - auto, 1:1, 16:9, 9:16, 4:3, 3:4. auto just lets the model pick.
• resolution - 1K or 2K, default 2K. Unlike Nano Banana Pro's native 4K, 2K is this model's ceiling.
• google_search - a boolean passed straight through to the API request; on Gemini image models that typically turns on search-grounded generation. Defaults to off.
• poll_timeout_seconds - default 600, min 30, max 3600. If a task legitimately runs long, bump it; if you want it to fail fast on a dead request, drop it.
• reference_image - optional IMAGE, single or batch, the input for image-to-image work.

Two outputs: images (the batch of generated images, wire it into SaveImage or your upscaling chain) and task_id (the Kie task string, useful for checking the job on Kie's side or debugging).

Installing it

If you're on ComfyUI Manager, just search the pack title ComfyUI-KieAPI. Otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/Song367/ComfyUI-KieAPI

Restart ComfyUI and you're done. No model downloads, no heavy dependencies - the requirements.txt is just aiohttp, numpy, Pillow and pydantic, all of which you almost certainly already have. What you do need is a Kie AI account and a pay-as-you-go API key; this is a paid remote service, not free compute.

Common issues

• "Kie API key is required" - you gave it nothing anywhere. Set the node input, KIE_API_KEY, or config.json.
• Task fails - the node raises with Kie's failCode/failMsg, so read the error. Model availability on resellers like this shifts (Kie's catalog has already shuffled models before), so an unexpected failure can mean the model name changed on their side.
• Slow or hung generation - raise poll_timeout_seconds before blaming the node; the default 10 minutes covers most runs, but 2K output with references can crawl.