Gemini 3 Pro Image

The class name is a mess - Gemini Pro Image (Nano Banana Pro) - and it's worth untangling because this is arguably the most interesting node in the pack. "Nano Banana" is Google's nickname for the Gemini image model, and this node routes Gemini 3 Pro Image through UCloud's Modelverse API. Google's image editing is genuinely strong - reference-image understanding, real instruction-following, coherent multi-image fusion - and it's absolutely something you cannot run on a home GPU. This node is the only way a lot of people will ever use it inside a graph.

What it is

A prompt-to-image generator with real editing chops. The headline difference from diffusion: you can attach up to five reference images (the node exposes image1 through image5) and it actually understands them - copy this subject's face onto that composition, keep the lighting of the first photo and the pose of the third. Diffusion models fake this with IP-Adapter and ControlNet stacks; Gemini does it natively. On top of that it can ground on live web search, which is what "search grounding" means here.

The inputs that matter

• client - required, from the Modelverse Client / Secret Client node.
• prompt - the instruction. Imperative and specific, like talking to a very literal art director.
• image1–image5 - optional reference images. This is the node's superpower; wire up to five and see how well it juggles them.
• aspect_ratio - auto or everything from 1:1 to 21:9.
• image_size - 1K, 2K, or 4K. Bigger is billed bigger. 1K is fine for iterating; use 2K/4K for final assets.
• mime_type - image/png or image/jpeg output.
• use_google_search - enable live search grounding for real-time info in the image. Great for "render a map of X" or "product shot with current packaging," but it can surprise you - the model now has internet.
• num_requests - batch count (1–10). The bill multiplier, again.

Output is a single image (IMAGE tensor). No URL/downstream step - it lands in the graph ready to preview or save.

Installing and keying up

The usual for this pack. ComfyUI Manager → search "ComfyUl-UCloud" → Install → restart, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ucloud/comfyui-plugin

restart ComfyUI. Light dependencies (requests, pillow, numpy, openai, GitPython), no model downloads - the model runs on UCloud's servers. Get a ModelVerse API key from the UCloud console and feed it via the Modelverse Secret Client node (secret name in the workflow, key in secrets.json) or a config.ini MODELVERSE_API_KEY=... entry.

Common issues

• "API key is not set" / 401 - the client node is empty or wrong. Always check there first.
• 4K feels wasteful for drafts - because it is. 1K for iteration, upscale tier only for finals.
• Search grounding gives inconsistent results - live grounding is a feature with a footgun: output depends on what's currently on the web. Disable it unless you need real-time info.
• Only five reference images wired when you expected more - the source mentions more, but the node exposes exactly five slots. That's the real interface.
• Costs - Gemini Pro Image through an API is metered and 2K/4K output plus search grounding is where the number gets interesting. Keep an eye on the console.

If you've ever assembled a five-model ControlNet stack to chase what this node does with five image inputs, you already know whether it's worth it.