Nano Banana 2.1

"Nano Banana" is Google's image model wearing a nickname that started as a leaderboard codename and became official branding. You can't download it, so a node like this is the only way it exists inside your graph - and the 2.1 entry in this pack is the general-purpose one: a prompt, an optional reference image, an aspect ratio, a resolution tier, and an output format. That's it. No seed, no sampler, no CFG.

Which is the first thing to internalise: there is no seed widget. Every run is a fresh roll. If you find an image you love, save it - you can't regenerate it. That single missing field changes how you work with this node compared to anything local.

How it works

The node submits a KIE Market job with the model nano-banana-2-1 and your request as its input, polls the task record until it succeeds, downloads the resulting image and hands it to ComfyUI as an IMAGE tensor - indistinguishable from a local sampler's output to everything downstream.

If you connect something to image_input, it's PNG-encoded and uploaded to KIE's file host before the call, then sent as the reference. The socket is treated as plural, so a batch of images goes up as multiple references rather than silently using the first one. This is the instruction-editing path: give it an image and a sentence, get the same subject in a new pose, outfit or scene. Identity survives because the model is looking at your reference the whole time instead of re-inventing it from a description - which is, per the KB, the 2026 answer to character consistency for stills.

The inputs

• prompt - required, up to 20,000 characters. Long prompts are explicitly supported, so use them: for edits, say what changes and what stays.
• image_input - optional. Empty for text-to-image; connected for edit and reference work.
• aspect_ratio - auto plus a big list (1:1, 2:3, 3:2, 16:9, 9:16, 4:5, and some extreme panorama shapes like 1:8 and 8:1). auto is fine for edits, since the reference usually dictates the frame.
• resolution - 1K, 2K, 4K. This is the cost lever: 4K is four times the pixels of 2K and you'll feel it in credits and wait time. Draft at 1K, re-render the keeper at 4K.
• output_format - png or jpg. PNG if it's going back into a graph; JPG if it's going to a client.
• timeout_seconds, callback_url - leave the callback empty.

Outputs: image, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left. all_urls_json matters more here than on most nodes - some providers return several variations, and this keeps every URL rather than just the one the node decoded.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
cd ComfyUI-KIE-Nodes-Next && pip install -r requirements.txt
# restart ComfyUI

Manager search: "KIE.ai Nodes Next". The pack needs requests, Pillow, numpy, PyYAML and nothing else - no model download, no VRAM. Key setup is one paste in Settings → KIE.ai Nodes Next → Connection; validated before saving, kept out of workflow JSON, masked in the UI.

Where people get burned

The missing seed is the big one. You will iterate by re-prompting, not by tweaking a number, and the good result is the one you can't get back. Save as you go.

Second: image edits drift. Each pass re-emits the whole frame, so faces slide a little per edit and the effect compounds across a chain. If you're holding a character across several images, get the mask or reference right rather than stacking edits.

Third, and unavoidably: your reference images leave the machine, the provider applies its own content policy, and billing is per call. Also note the node is not an output node - if image goes nowhere, nothing runs and nothing gets charged. And use the right-click KIE Next · Check inputs & model controls menu to see the documented limits without uploading or spending anything; going straight to a paid call to discover an aspect-ratio rule is an expensive way to learn.