JILIGULU Nano Banana Pro

This is the node you reach for when "good enough" isn't the assignment. JILIGULU Nano Banana Pro is the premium tier of a cloud API-client pack: it sends your prompt and up to eight reference images to jiligulu.art, which resells Google's Nano Banana (Gemini Image) family, and returns a ComfyUI IMAGE. The service pitches Pro at exactly the jobs local models choke on - posters, product shots, and visual design where text rendering and subject consistency have to actually hold up.

The Nano Banana lineage matters here, because Pro is the flagship of the family: 4K-native generation, thinking-style reasoning before it draws, and multi-language text rendering are the things the community actually talks about when it compares closed image models. The KB's verdict on the family is worth keeping in mind too - powerful but "unimaginative," and aggressively filtered. Pro buys you polish, not creative freedom.

One honest caveat: don't assume this API's numbers match Google's official specs. JILIGULU caps Pro at 8 reference images and up to 4K; the node enforces those caps locally, before any request ships. Go by what the node tells you, not the marketing.

How it actually works

Same boring, dependable machinery as the rest of the pack. It fetches the model config from GET /api/v1/models/nano-banana-pro/config, uploads each reference image as PNG, posts a run job, polls the task endpoint every 5 seconds for up to 15 minutes, then downloads the result into an IMAGE tensor. No local inference, no model downloads, and - refreshingly - no new Python dependencies. Just the standard library plus numpy/torch/PIL that ComfyUI already has.

The inputs that matter

• prompt - required, and on a model like this it's where the thinking happens. Empty prompt = instant local error.
• api_key - in the node or via JILIGULU_API_KEY set before launch.
• aspect_ratio - auto plus ten ratios (1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 5:4, 4:5, 21:9).
• image_size - 1K, 2K, or 4K. 4K is Pro's whole point; it's also where the credits go.
• image / image_2 … image_8 - reference inputs, capped at 8. Use them for subject consistency: a product photo, a character sheet, a brand look you want held across generations.

Output: image (IMAGE). From there it can go anywhere - save node, upscaler (though 4K output usually doesn't need one), or further image-to-image passes downstream.

Installing it

ComfyUI Manager → search "JILIGULU Nano Banana" → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/XYQC-LH/ComfyUI-JILIGULU-Nano-Banana

Restart and you're live. No requirements install, no weights to place in models/ - this pack is genuinely low-friction on the dependency front.

Common issues

• The cost of 4K adds up. A 4K Pro run is a bigger cloud job than a 1K one, and every run spends credits with JILIGULU. Draft at 1K, commit to 4K only for the final pass.
• Chinese error messages. The author raises errors in Chinese; they're accurate once translated. The usual suspects are a missing key and reference-count overflows.
• Batch frames count individually. A 4-frame batch into one input consumes four of your eight references. Watch it if you're wiring up an image feed rather than a single Load Image.
• Trust and privacy. This pack is brand new and "jiligulu" has essentially no footprint in the English community at the time of writing. Your prompts and reference images leave your machine for a Chinese aggregator; the underlying model is Google's, with heavy filtering and an invisible SynthID watermark baked into every image. Test with something trivial first, and don't send it anything you wouldn't hand to a stranger's API.

If your workflow needs commercial-grade output in the loop, Pro is the one you'd actually keep wired in. Just budget for it.