🍌 GrsAI Nano Banana 2 - Text/Image

The middle child of the banana family

🍌 GrsAI Nano Banana 2 wraps Google's Nano Banana 2 - the Gemini 3.1 Flash image model, Google's speed-and-quality compromise that landed in February 2026. In the ComfyUI-GrsAI pack it sits between the budget Nano Banana node and the flagship Pro node: faster than Pro, smarter than the base model, and priced accordingly (roughly ¥0.06 per image, about a cent). It's the node you reach for when the cheap banana is too dumb and the Pro is overkill.

Where it earns its keep over the base node: you get image_size control (1K/2K/4K), ten reference-image inputs instead of six, and a wider aspect-ratio list that includes ultra-wides like 4:1. The model is also a genuine step up in prompt adherence and editing quality while keeping Nano Banana's fast turnaround. Community verdict on the family still applies - Google's filters are aggressive and opaque, and every image carries the SynthID watermark - but as a "quality for the price" pick, this is the one.

How it works

Identical machinery to the rest of the pack, which is part of its charm. The node reads your GrsAI key from the apikey field, sends your prompt plus any reference images (as base64 PNG data URIs) to the GrsAI relay at https://grsai.dakka.com.cn, and downloads the result back into an IMAGE tensor. No images wired in = text-to-image; one image = editing; a handful = multi-image fusion. num_images spawns that many concurrent paid requests - each is its own API call, so batch of 4 means four parallel bills.

The inputs that matter

• apikey - your GrsAI key in the field on the node. That's what the code actually reads.
• prompt - the text. Nano Banana 2 handles detailed prompts and in-image text well.
• model - nano-banana-2 (default) plus the reseller SKUs nano-banana-2-cl and nano-banana-2-cl-4k. The differences are GrsAI's own tiers; stick with the default unless the others show up in their docs.
• num_images - 1–12 concurrent requests.
• image_size - 1K / 2K / 4K, default 1K. This is the big differentiator from the base Nano Banana node, and yes, it's real here: the code's support list includes nano-banana-2 (the README's older claim that only Pro honors imageSize is stale).
• aspect_ratio - auto plus fourteen ratios including 4:1, 21:9 and friends.
• image_1 … image_10 - up to ten reference images.

Outputs: image (tensor → SaveImage) and status (STRING summarizing model, image size, reference count and success/failure counts).

Installing it

Same one-time pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/31702160136/ComfyUI-GrsAI.git
cd ComfyUI-GrsAI
pip install -r requirements.txt

Restart ComfyUI and the whole "🍌" family appears under the GrsAI menu. Dependencies are light - requests, httpx, python-dotenv, Pillow, torch - with no model downloads. Windows portable users: .\python_embeded\python.exe -m pip install --force-reinstall -r .\ComfyUI\custom_nodes\ComfyUI-GrsAI\requirements.txt. If ComfyUI Manager blocks the git install, the README's documented fix is setting security_level = weak in the Manager config.ini.

Where people get burned

• The key lives on the node. The README's .env / GRSAI_API_KEY flow serves the pack's Flux nodes; this node reads the apikey widget. Paste the key there. The placeholder Chinese text it ships with will 401.
• Watch the image_size bill. 2K and 4K cost more per image - the README lists prices per tier. If you're iterating, leave it at 1K; jump to 4K only for finals.
• It always re-runs. IS_CHANGED is pinned to fire every queue, so each run is a paid generation. Don't spam Queue and wonder why the credits are gone.
• Failures are silent-ish. A failed call returns a 1×1 black tensor and a 失败: ... status string instead of throwing. If your SaveImage writes black frames, read the status output - and if everything fails at once, check your GrsAI balance.

For most "I want Google's model but not its full API" workflows, this is the node I'd reach for. The Pro is better, sure - but it's also 50% pricier and you'll only notice the difference on output you actually care about.