JILIGULU Nano Banana 2

Nano Banana 2 is the newest member of Google's Gemini-image family - a speed-and-quality hybrid that shipped in early 2026 - and JILIGULU Nano Banana 2 is the cloud-API node that brings it into ComfyUI. Its superpower is reference handling: it accepts up to fourteen reference images, more than any other node in this pack, which makes it the one you want for multi-angle character sheets and multi-round visual edits where the model has to hold a subject together across a pile of inputs.

The trade is in the resolution dropdown. This node offers 1K and 2K only - no 4K. If you need native 4K output you want the Pro node; if you need reference bulk, this is it. The two tiers are complementary, not competing.

Why would you actually reach for this? The community's favorite loop is generating character sheets with Nano Banana and feeding them into LoRA training. A fourteen-reference node is built for exactly that: give it a front view, back view, outfits, close-ups, and it holds the subject consistent through the whole set. That's the workflow this node exists for.

How it actually works

It's the same base-class machinery as the rest of the JILIGULU pack, targeting nano-banana-2 on jiligulu.art: fetch the model config, upload every reference image as PNG, post a run job, poll the task endpoint every 5 seconds for up to 15 minutes, then download the result into a ComfyUI IMAGE. The reference upload is per image, so a fourteen-reference run is fourteen separate uploads - give it a moment before the actual generation even starts. No local GPU work, no model files, and no new Python dependencies (stdlib urllib plus numpy/torch/PIL, all already present in ComfyUI).

The inputs that matter

• prompt - required; empty prompts are rejected locally.
• api_key - in the node, or set JILIGULU_API_KEY before launching ComfyUI.
• 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 or 2K. Fourteen references at 2K is the realistic ceiling of what this tier is for.
• image / image_2 … image_14 - the star of the show. Cap of 14, enforced locally before any request leaves your machine.

Output: image (IMAGE). Save it, preview it, or pipe it into further graph steps.

Installing it

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

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

Restart ComfyUI. There's no requirements.txt to fight and no model to download - this pack's whole appeal is that the dependency story is trivial.

Common issues

• Fourteen references is also fourteen uploads and one big bill. Every run spends credits with JILIGULU. The reference-heavy strength of this node is exactly where the meter runs. Draft with fewer refs, then commit to the full set for the final pass.
• Batch frames count as individual references. Feed a 5-frame batch in and that's five of your fourteen slots gone. If your reference inputs come from a video or a batch loader, keep an eye on it.
• Chinese error messages. The author's errors are in Chinese and they're accurate once translated - missing key, upload failure, task timeout, that family of problems.
• It's an unknown, paid, third-party API. The pack is brand new and "jiligulu" has essentially zero presence in the English community at the time of writing. Your prompts and reference images leave your machine for a Chinese aggregator, and the underlying Google model ships with heavy content filtering and an invisible SynthID watermark. Test cheap, read the terms, and don't feed it anything sensitive.

For building reference libraries and character sets, it's genuinely the strongest node in the pack. Just remember the 2K ceiling before you plan a 4K pipeline around it.