Gemini 3 Pro Image

Gemini 3 Pro Image is the money node of this pack: it sends a prompt and up to ten reference images to Google's Gemini image models and gets back a real ComfyUI IMAGE tensor you can preview, save, or pipe into the rest of your graph. Pick nano-banana-pro and you're running Gemini 3 Pro Image Preview - the "Nano Banana Pro" model, native 4K output and a built-in thinking step. nano-banana gives you Gemini 2.5 Flash Image, nano-banana-2 the newer Gemini 3.1 Flash Image.

Here's the honest tradeoff up front. Nothing downloads, nothing runs on your GPU - it's all Google's hardware, billed per image at roughly $0.04–$0.24 depending on resolution, and the community's ballpark of ~$0.25/image still holds for heavier calls. That makes it a poor fit for batch experimenting, and it's worth being blunt: Nano Banana is aggressively censored. It's powerful, and also "unimaginative" - famous IP and celebrity prompts bounce, and every output carries an invisible SynthID watermark. Reach for it when you need what it's actually good at: multi-image editing with strong identity consistency, crisp text-in-image, and a model with real world knowledge behind the prompt. Keep your local stack for everything else.

How it works

The node builds a Gemini generate_content request with your prompt, any reference images (sent as PNG), and an ImageConfig carrying aspect ratio and size. The model aliases have a fallback chain: nano-banana tries 2.5 Flash Image then falls back to 3.1 Flash Image Preview, nano-banana-2 goes the other way - handy when one endpoint isn't available. API keys rotate automatically on quota errors, same as the text node. You get two outputs: the image and a text string of whatever the model said alongside it.

Inputs that matter

• model - the three Nano Banana aliases; the README's mapping is what the API actually sees
• prompt - what to generate or edit
• aspect_ratio - auto plus 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9, 4:1, 1:4, 8:1, 1:8; auto picks the closest match to your first reference image
• image_size - 0.5K, 1K, 2K, or 4K (0.5K is sent as 512 and is only supported by the 3.1 Flash models)
• thinking_level - default, minimal, or high; only the 3.1 Flash models accept explicit levels

Then reference_image through reference_image_10 for editing, and seed for reproducibility. The api_key input is optional if you've set config.json or GEMINI_API_KEY.

The traps, learned from the source

Two things bite people. First, thinking_level of minimal or high on nano-banana-pro doesn't work - Gemini 3 Pro Image Preview uses its own default dynamic thinking and the API won't let you override it, so the node errors on purpose rather than silently ignoring you. Leave it on default for Pro. Second, don't reach for 0.5K with nano-banana - only the 3.1 Flash models take the 512 size, and the node will skip to compatible candidates or fail. thinking_level behaves the same way. Both checks are enforced in code with clear error messages, so you'll know immediately.

Install is the same as the whole pack - ComfyUI Manager (search "gemini-3-simply-comfyui") or clone, copy config.example.json to config.json, restart. If you hit [SSL: WRONG_VERSION_NUMBER]-style errors on the Windows portable build, it's almost always a proxy or HTTPS-inspecting VPN interfering with the API call, not a node bug. And if you get a "model did not return an image" error, that's usually the safety filter refusing the content - the model's text reply will tell you why.