Nano Banana 2
14 reference images plus live web search
- images
- image
Nano Banana 2 is Google's Gemini 3.1 Flash image model - the February 2026 speed-quality hybrid in the line that started with the original Nano Banana. This node runs it through the Kie.ai API, which matters for two reasons: you get it inside a ComfyUI graph without writing any API code, and you get it at Kie's marketplace rates instead of paying per-call through another gateway.
What makes Nano Banana 2 genuinely interesting versus its Pro sibling is the input side: up to 14 reference images plus an optional Google web-search grounding toggle. That combination turns the node into a "remix this vibe with current information" tool. Point it at a few product shots and a style reference, flip web search on for something time-sensitive, and the model has a lot more to work with than a plain text prompt.
How it works
Like every node in this pack, it's a thin async client for the Kie.ai job API. Your prompt and any reference images are uploaded, the task is submitted, and the node polls until the image is ready, then decodes it into a normal ComfyUI IMAGE tensor (BHWC float32 0–1) - which means every downstream ComfyUI node, save, or upscaler treats it as a native image. Nothing renders locally; the cost is Kie credits, not GPU hours.
Inputs that matter
- prompt - the only required input. Multiline, and the model is Gemini-flavored, so it rewards plain descriptive language.
- images - optional reference images, up to 14. Pass a batch (e.g. from LoadImage or a prior gen) and each is used as a reference.
- google_search - the web-grounding switch, default off. Flip it on when the result should incorporate real, current information.
- aspect_ratio - defaults to
auto, with the usual presets available. - resolution -
1K,2K, or4K. The 4K step is where the credit cost jumps. - output_format -
jpg(default) orpng. JPEG is cheaper on bytes; choose PNG if you're chaining into editing workflows.
Output: a single image (IMAGE) that wires into SaveImage or any image consumer.
Install and key setup
No model downloads - this is an API node. Install from ComfyUI Manager (search "ComfyUI-Kie-API") or:
cd ComfyUI/custom_nodes
git clone https://github.com/gateway/ComfyUI-Kie-API
Restart, then create config/kie_key.txt from config/kie_key.example.txt and paste in your Kie.ai key. Confirm with the pack's "Get Remaining Credits" node before the first real run.
The honest caveats
- It's Google's model, behind a Google-shaped filter. The KB on closed image models documents Nano Banana's aggressive IMAGE_SAFETY filtering and its strict stance on famous IP. Don't expect the open-model level of control you get from a local checkpoint - Nano Banana 2 is a hosted service and it behaves like one.
- Web search is a feature and a cost. Grounding on live info is genuinely handy, but every search-enabled call is doing more work. Use it deliberately, not as a default.
- 14 references is a lot of upload. If your batch is huge, expect the job to take a bit longer on the upload side. The 2000s timeout has headroom, but a 4K multi-reference run is a sit-and-wait, not a snappy local gen.
For storyboard and reference-heavy work, this is the more flexible of the two Nano Banana nodes in the pack. The Pro node is the quality flagship; this one is the utility player with the deeper input bench.
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| prompt | STRING | — | |
| imagesopt | IMAGE | — | |
| google_searchopt | BOOLEAN | false | — |
| aspect_ratioopt | COMBO | auto | 15 options: 1:1, 1:4, 1:8, 2:3, 3:2, 3:4, +9 |
| resolutionopt | COMBO | 1K | 3 options: 1K, 2K, 4K |
| output_formatopt | COMBO | jpg | 2 options: jpg, png |
| logopt | BOOLEAN | true | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | — |