CatsAPI Nano Banana 2

Google's Nano Banana line is the most-covered closed image family in the whole API-node ecosystem - "nano banana" saturated the reddit corpus cap within months, and there's an entire reseller market around it. Nano Banana 2 is the middle child: the Gemini 3.1 Flash–based speed/quality hybrid that lands between the consumer base model and the flagship Pro. You can't download any of them; you call them. CatsAPINanoBanana2 is one of seven nodes in the ComfyUI_Catsapi pack that do exactly that against catsapi.com, and it's the one I'd reach for as the pack's best everyday draft button.

The inputs

• prompt - multiline text.
• resolution - 512px, 1K, 2K, or 4K. That 512px option is the quiet hero: a cheap, fast draft tier that makes iteration painless before you spend on a full-size render.
• aspect_ratio - fourteen presets, and here's where it gets fun: alongside the normal 1:1/16:9/9:16 set, Nano Banana 2 accepts extreme strips like 1:4, 1:8, 4:1, and 8:1. That's banner material, and it's the model's own capability - the node only lists what the API genuinely accepts.
• num_images - 1 to 4 per run.
• max_coins - spending cap; 0 = uncapped.
• reference_image (optional) - up to four IMAGE tensors for editing/composition.
• api_key_override (optional) - leave blank locally.

Outputs: the standard images tensor, file_paths (JSON), cost_coins, task_id, metadata. Files download to ComfyUI/output/catsapi/ and become a normal tensor, so you can chain a local upscaler - which pairs beautifully with that cheap 512px draft tier: draft tiny, then scale locally.

Why this one, not Pro

Nano Banana 2 is the speed play. Pro is the flagship with "thinking mode" and 4K native output, but it's also pricier and has no 512px tier. If you're doing iteration - design concepts, storyboard looks, dozens of variants - the hybrid's lower cost per image plus the draft resolution is the sane choice. The reference-image count (four) also makes it a solid character-consistency workhorse.

How it works

Same skeleton across the pack: preview cost, check max_coins, submit, poll every three seconds (15-min timeout), download with browser headers and a curl fallback. Nothing runs locally; the node is a thin HTTP client over Google's model as resold by catsapi.com.

Install and the key

cd ComfyUI/custom_nodes
git clone https://github.com/maodeyu180/ComfyUI_Catsapi.git
# restart ComfyUI

No extra Python dependencies - urllib plus numpy/PIL/torch that ComfyUI already has. You need a paid CatsAPI key from catsapi.com (starts with cats-):

export CATSAPI_API_KEY=cats-your-key
python main.py

Shell env not inherited? Use a .env or ~/.catsapi.env file; the node reads those too.

Where people get burned

The usual API-node trio: your prompt and reference images leave the machine, api_key_override gets serialized into saved workflow JSON (don't share it filled in), and the provider is a young Chinese reseller with no English-community footprint - price in cat-coins, verified by your own drafts. Also remember the model's own filter: Google's moderation rides along no matter which reseller you route through. The 8:1 banner ratios are fun until you render a dozen 4K ones on a whim; that's what 512px and max_coins are for.