JILIGULU Nano Banana

The name suggests a chunky local checkpoint. It isn't. JILIGULU Nano Banana is a thin API client node: it takes your prompt (and optionally up to ten reference images) and ships them to a cloud service called JILIGULU (jiligulu.art), which resells Google's "Nano Banana" - the Gemini-native image model family - through a simple HTTP API. Your GPU does none of the work. Your wallet does some of it, and the images do come back as a normal ComfyUI IMAGE.

Why reach for it? Nano Banana is genuinely good at things local models still fumble: clean text rendering, subject consistency across a few reference images, and reference-based editing. If you've ever wanted Gemini-grade output piped into the rest of your graph - upscale it, img2img it, composite it - this is the plug. The community does exactly that: generating character sheets and multi-angle reference sets with Nano Banana and feeding them into LoRA training runs. That's the killer workflow here, and it's why the reference-image inputs matter so much.

How it actually works

Open the source and it's boring in the best way - no requirements.txt, no model files, just Python's standard urllib plus numpy/torch/PIL, which ComfyUI already bundles. The request chain:

1. GET /api/v1/models/nano-banana/config - pull the server's defaults for the model
2. Upload each reference image as PNG via POST /api/v1/assets/upload
3. POST /api/v1/models/nano-banana/run with your prompt, aspect ratio, and size
4. Poll GET /api/v1/tasks/{task_id} every 5 seconds, up to 15 minutes
5. Download the result and convert it to an IMAGE tensor

So when you hit Run, the node isn't generating anything locally - it's queueing a job on a remote machine and babysitting it.

The inputs that matter

• prompt - the only thing that always matters. Empty prompt is a hard local error.
• api_key - the node takes it directly, or you can set JILIGULU_API_KEY as an environment variable before launching ComfyUI. If neither exists, you get an error before anything leaves your machine.
• aspect_ratio - auto plus the usual 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 5:4, 4:5, and 21:9.
• image_size - 1K, 2K, or 4K for this node.
• image / image_2 … image_10 - optional reference images. This node caps at 10, enforced locally so a bad request never reaches the backend.

One output: image (IMAGE). Wire it into a PreviewImage or SaveImage, or anywhere an IMAGE is expected.

A real gotcha: the node flattens batch frames. Feed it a 4-frame batch from one input and that counts as four references. The cap is per frame, not per input socket.

Installing it

Easiest: ComfyUI Manager → search "JILIGULU Nano Banana" → install → restart. Or the manual route:

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

Then restart ComfyUI. No pip step, no model downloads. That's the whole install, and it's the pack's main virtue.

Common issues

• "缺少 api_key" and other Chinese errors. The author raises errors in Chinese. Translate them - they're usually accurate (missing key, upload failed, task timed out).
• A generation can take a while. The node polls for up to 15 minutes per job. This is not a fast local sampler.
• It costs money. Every run spends credits on JILIGULU's side. Don't fire off a ten-reference batch to test a typo'd prompt.
• It's an obscure service. At the time of writing, "jiligulu" has effectively zero footprint in the English community, and this pack is brand new. Treat it like any small third-party API: your prompts and images leave your machine for a Chinese aggregator, and the underlying Google model is heavily filtered and stamps an invisible SynthID watermark. Read the terms, test with a trivial prompt first, and don't put confidential material through it.

The reference-editing and character-sheet workflows are genuinely compelling. Just know exactly what you're plugging into your graph before you make it a permanent fixture.