EvoLink Nano Banana 2 Lite (Official)

Every closed-model family gets a budget tier eventually, and Nano Banana's is the Lite: Google's Gemini 3.1 Flash Lite Image, the cheapest way to get Banana-quality generation without paying for the flagship. EvoLink Nano Banana 2 Lite (Official) is that tier as a node. The trade is simple - output is fixed at 1K, always - but if your job is batch work, prompt iteration, or debugging a workflow before you spend real credits on the good stuff, that's exactly the trade you want.

What it is

Text-to-image, image-to-image, and image editing, same as its big sibling, minus the resolution ladder and the web-search knob. The params are trimmed to match:

• prompt (required) - the description or edit instruction, Chinese welcome.
• size - auto or ratio presets, including the extreme 1:8 / 8:1. The model picks framing within the fixed 1K output.
• thinking_level - auto / min / high, default auto. Same "reasons before it paints" knob as the full Nano Banana 2, if a bit redundant at budget tier.
• image1–image4 - reference image inputs. Plug a Load Image into image1 and you're doing img2img/editing; your local image uploads via EvoLink's channel and is auto-deleted after 72 hours. Up to 14 references when you add image_urls_text.

Outputs are the pack standard: image (downloaded tensor - connect Preview/Save Image or you get nothing visible), task_id, status_info, result_urls (24-hour expiry), response_json.

Install and key

It's the same EvoLink install for all of these nodes, so if you already have the pack, skip this. ComfyUI Manager → search EvoLink → install → full restart (close the console window). Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/deeplearning-goethe/comfyui-evolink

Dependency is just requests. No models, no GPU. Key from evolink.ai/dashboard/keys (sk-), pasted once into api_key; it saves locally and can stay blank after, so it never ships in a shared workflow JSON.

Where it fits

Use it as the "is my prompt even good" tier: validate a workflow's wiring and prompt at 1K, then swap in Nano Banana 2 or Pro for the final pass. That's the pattern that keeps your bill honest. The fixed 1K also means you're not tempted to run every draft at 4K just because the dropdown is there.

Downsides worth naming: 1K is fine for a hero image and weak for print or large crops, and there's no 4K escape hatch here - if a draft comes out great and you want it big, you re-run on a higher tier. Also, as with every EvoLink image node, remember it's metered and your prompts leave the machine. For a closed model you can't run locally that's the deal; the Lite tier is just the version of the deal that hurts least while you iterate.