EvoLink Nano Banana 2 (Official)

Nano Banana is the closed image model this community actually searches for. It's Google's Gemini 3.1 Flash Image dressed in a fruit costume, and "nano banana" saturates Reddit to the point where it outpaces every other closed image model's discussion. You can't download it - Google keeps it firmly API-only - so the only way it gets to sit in your ComfyUI graph next to your local upscaler and masking stack is a node like this one.

EvoLink Nano Banana 2 (Official) is the mainstream tier: the speed-quality hybrid that replaced the original Nano Banana as the default. Behind the scenes the node maps nano-banana-2 to Google's gemini-3.1-flash-image-preview endpoint, so you're getting the real thing, not a lookalike.

What it does

Text-to-image, image-to-image, and image editing in one node. The img2img part is the killer feature: drop a Load Image node into image1 and you're editing - your local image uploads through EvoLink's official channel (auto-deleted after 72 hours, no third-party image host, no imgbb nonsense). There are four image slots (image1–image4) plus an image_urls_text box for public URLs, with up to 14 references total.

The settings that matter:

• prompt (required) - describe or instruct. Chinese works, roughly 2000 tokens of room.
• size - auto default, or ratios from 1:1 out to the extremes (1:8 / 8:1). If you need a banner or a vertical story panel, this is one of the few image nodes that lets the model actually do extreme aspect ratios.
• quality - 0.5K / 1K / 2K / 4K, default 2K. Note this means something different from the GPT Image quality knob - here it's the output resolution tier. 4K costs extra.
• thinking_level - auto / min / high. This is the "reasons before it paints" mode, Google's thinking knob. auto is the sane default; high is slower and for the hard prompts.
• web_search - optional web search to ground the generation. May add cost, so don't leave it on by accident.

Outputs: image (already downloaded into a tensor - connect Preview Image or Save Image or you'll see nothing), task_id, status_info, result_urls (24-hour expiry), response_json.

Install and key

Standard EvoLink install: ComfyUI Manager → search EvoLink → install → full restart (close the console, not just refresh the browser). Manual alternative:

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

Only dependency is requests; nothing downloads, nothing runs locally, no GPU needed. Get a key from evolink.ai/dashboard/keys (sk-), paste it into api_key once - it's saved locally, so leave it blank afterward and it won't leak into a shared workflow.

Gotchas

The classic: task completes, no image. That's the missing Preview/Save Image connection - reconnect and re-run; the cache means no double charge. Reference images with real people in them: the tooltip notes the model is happiest with 4–5 max for photo-real refs, so don't feed it a dozen portraits and expect coherent faces. And remember what you signed up for with any cloud image node: per-call cost and your images leaving the machine. For a model you can't run locally, that's the price of admission - just don't leave web_search on and run a thousand-image batch.