google / nano-banana-2

Here's the thing nobody else in this pack can do: enable_web_search. CivitaiImageGenGoogleNanoBanana2 is the Gemini 3.1 Flash–based speed-quality hybrid (the sequel to the model everyone knows as Nano Banana), and its headline feature is that it can go look things up before it draws. Ask for a landmark, a building, a current event, and instead of hallucinating a plausible-looking fake, the model has a shot at generating something close to what actually exists.

It's one of the ~160 nodes in the official civitai/civitai-comfy-nodes pack (engine google, model nano-banana-2). Prompt in, IMAGE out, nothing runs on your machine - the job executes on Civitai's cloud and you pay in Buzz.

How it works. Standard pack mechanics: build a payload, POST to orchestration.civitai.com, long-poll until terminal, return the image. What's unusual is the input list. aspect_ratio covers the wide modern set (21:9 down to 9:16, default 1:1). resolution goes 1K/2K/4K - a real step up from the base Gemini node that has no size control at all. num_images (1-4) and seed are there, plus an images input, which means this "2" model also does reference-image editing, not just creation. And then the pair that makes it special: enable_web_search and enable_google_search (both default off). The distinction is subtle - one searches the wider web, the other Google's own index - and honestly, for most prompts the web toggle is the one to flip. The model then researches your subject and incorporates what it finds into the generation.

Inputs that matter. prompt (required) is the start. If you're generating something fact-anchored - a specific landmark, a product, a recognizable person-adjacent scene - flip enable_web_search on and see if the results stop being "imagined." resolution is the quality dial: 4K costs more Buzz than 1K and is the model's native best. And images lets you feed references for style or subject consistency, which is how you get the character-consistency workflows people actually use these models for.

Install. The whole pack installs at once - ComfyUI Manager → search Civitai Comfy Nodes (publisher civitai) → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

No model downloads; the model lives in Google's cloud behind Civitai's API.

Auth and gotchas. Credentials from a Civitai Auth node, then CIVITAI_API_TOKEN, then stored OAuth. Error hints to recognize: 402 = insufficient Buzz, 401 = token expired, 429 = rate limit. 30-minute job timeout, early-preview pack so shapes can change.

The honest caveats: Nano Banana models are famously "powerful, unimaginative," and heavily censored - Google's filter is baked in, Civitai's is on top, and no node setting bypasses either. Web search also sends your prompt out for an extra lookup, so the "your data leaves the machine" asterisk grows a little larger. But for the one job this node uniquely nails - factual image generation from text - it's the only member of the pack that even tries.