Gemini Image Generate

Gemini Image Generate puts Google's Gemini 3 Pro Image - the model behind the "Nano Banana Pro" branding, the one that made the KB's closed-source-models essay sit up - into a ComfyUI graph as a prompt-to-IMAGE node. It's the newest member of this pack (added January 2026 per the README), and like everything else in ComfyUI-MidjourneyHub it's a thin client for the 云雾 (Yunwu) relay: no local model, no GPU load, pay per image. What you're buying is Google's headline image quality - strong text rendering, multi-language support, and that "reasons about the prompt before drawing" behavior - without maintaining a Gemini API key of your own.

How it works

The node builds a standard Gemini generateContent request and POSTs it to {api_url}/v1beta/models/gemini-3-pro-image-preview:generateContent, with responseModalities set to ["TEXT", "IMAGE"]. The response's inline image data arrives base64-encoded; the node decodes it, converts it to a 0–1 float tensor, and returns it. That's the whole pipeline - one blocking HTTP call with no polling, because Gemini's generateContent returns the finished image (or images) in the response.

Inputs

• prompt - required, multiline.
• aspect_ratio - 1:1 (default), 9:16, 16:9, 4:3, or 3:4. Note this is one field for both orientation and ratio - there's no separate width/height.
• image_size - 1K (default) or 2K. The 2K option is where Gemini's high-res advantage shows; it'll cost you more per image.

One output: IMAGE, wired straight to a preview or save node.

Installing

ComfyUI Manager (search "ComfyUI-MidjourneyHub") or:

cd ComfyUI/custom_nodes
git clone https://github.com/jiaqianjing/ComfyUI-MidjourneyHub
# restart ComfyUI

Your key goes in config.ini under [GEMINI_API] - separate section from the Midjourney and GPT keys, and also pre-filled with the author's, so replace it before spending credits. Dependencies are the pack's usual light set (aiohttp, openai, nest_asyncio); no model files to download.

Where people get burned

Same relay caveats as the pack's other nodes: Chinese-language console, pricing groups that can refuse certain models ("无可用渠道"), and an api_url the author warns may change without notice. Gemini-specific expectations: Google's image models are heavily filtered, so don't go shopping for uncensored output here - the KB notes ImageFX is "powerful, unimaginative, massive hidden biasing" territory. Use it for what it's best at: promptable, high-res, text-accurate generation inside a ComfyUI workflow.