Gemini-Nano-2-pro图片编辑

Nano Banana Pro is one of the most-hyped image models of the mid-2020s - Google's flagship closed generator, 4K native output, genuinely strong instruction-following, and text rendering that most open models still can't touch. You cannot download it. There are no weights. The only door is an API, and this node (display name "Gemini-Nano-2-pro图片编辑") is that door inside ComfyUI: it wraps MojieAI's hosted nano-banana-pro endpoint so you get the model in your graph without ever creating a Google Cloud project, a billing account, or a Vertex AI anything.

It's also the pack's most flexible image tool. The prompt field takes natural language - describe the edit or the scene in plain terms, not SD syntax. Chinese input works, and unlike the pack's older NanoBanana node this model handles Chinese output too (the KB's closed-model notes flag that as a genuine Nano Banana Pro selling point). Wire in input_images and it becomes an editor - change the background, restyle an object, swap a color - rather than a plain generator.

How it works

Your prompt (and optionally your images) get base64'd and POSTed to Mojie's endpoint under the nano-banana-pro model; the returned image URL is downloaded and dropped back as an IMAGE tensor. The pack adds two bits of housekeeping: reference images are auto-compressed so their long edge doesn't exceed 1280px (a sensible cap for upload cost - know that your refs get downscaled, so feed sharp sources), and if you leave aspect_ratio on auto with no input image, it silently becomes 1:1 rather than guessing.

Inputs that matter

• prompt - required, and it does the heavy lifting. Keep it short and specific; this model rewards plain language over keyword soup.
• resolution - 1K, 2K, or 4K. This is the biggest cost driver: 4K is the model's flagship output and it's priced accordingly.
• aspect_ratio - auto plus nine presets (16:9 down to 9:16). Pick one and it's respected; leave auto and the model decides.
• num_images - 1 or 2. Two per call if you want a pick.
• is_translation - a translation toggle for when your Chinese prompt isn't landing; flip it and the backend pre-translates before generating.
• input_images - optional reference/editing images.
• seed - reproducible-ish, normal diffusion behavior.

Output is output, an IMAGE batch (batch size = num_images).

Installing it

Same pack ritual, twice-told because it's the same for every node here: ComfyUI Manager search "mojieapi_party", or git clone into custom_nodes; grab a key from mojieaigc.com into config.ini; restart. requirements.txt is light (the notable extra is opencv-python), and there are no model downloads - the whole point of an API wrapper.

The honest tradeoffs

This is where I stop selling. Nano Banana Pro is a closed, metered, filtered model. Per-image cost adds up fast, and you're paying a reseller markup on top of Google's own $0.039–0.24/image pricing. Everything you send - prompts and reference images - leaves your machine and lands on a server you don't control, subject to Google's moderation. The KB's API-nodes essay has the canonical community take: API nodes are the right tool for a model you cannot run, and the wrong default for one you can. For a closed flagship like Nano Banana Pro, there is no local option, so the equation is simple - but know the price before you wire it into every workflow. The red-image-with-text error channel (401 key, 403 balance) is your friend here; watch it.