🍌 智绘_banana2_向量引擎api (增强版)

Another way to run a closed Google image model inside ComfyUI without renting a GPU: ZH_BananaVectorAPI (🍌 智绘_banana2_向量引擎api 增强版) is the 智绘灵箱 pack's wrapper for the VectorEngine endpoint. Same shape as the pack's ZH_BananaTutuPort - a key, a prompt, an HTTP call, an image back - but pointed at https://api.vectorengine.ai by default, with a slightly different feature set and an "Auto (From Image)" aspect-ratio mode that the other node doesn't have. Its source header calls it the "Gemini 3 Pro (香蕉) V17 增强版," which tells you exactly which model family you're proxying: the closed nano-banana/Gemini image line.

How it works

Mechanically it's boring, which is the point: a requests.Session (reused across runs so you don't pay connection setup every call - that's the "增强版" performance bit) POSTs your JSON payload to base_url, gets back a base64 image, decodes it, and hands it to you as an IMAGE tensor. Your api_key is the credential; prompt is the instruction; seed gives you reproducibility; and up to four optional ref_image1–ref_image4 sockets send reference images along for editing or subject-consistency jobs.

The differences from the TutuPort node are the useful ones:

• aspect_ratio includes Auto (From Image) - when you feed reference images, the output aspect ratio follows the reference instead of a fixed list. That's genuinely handy for edits where the input geometry should dictate the output.
• resolution (1K default, or 2K/4K) replaces the separate image_size control.
• image_quality (default 90) is the JPEG compression dial - again, file-size control, not model fidelity.

Like its sibling, it's an output node (OUTPUT_NODE = True) with three sockets: 生成图片 (IMAGE), 详细信息 (STRING) raw detail, and 智绘显示 (STRING) - the pack's formatted status panel.

Install

Part of the 智绘灵箱 (ComfyUI-ZhiHui) pack:

cd ComfyUI/custom_nodes
git clone https://github.com/zhuyungen/ComfyUI-ZhiHui.git

Restart ComfyUI (or ComfyUI Manager, "智绘灵箱" / "ComfyUI-ZhiHui"). Runtime deps are just torch/numpy/requests.

Where people get burned

• VectorEngine is a reseller, not the model maker. You're trusting a third party to proxy Google's model and to keep their endpoint, rate limits, and billing sane. The KB's external-api-nodes doc covers this ecosystem: cheap, region-friendly, and one step removed from first-party. Verify your key works on the provider's own dashboard before blaming the node.
• The default base_url may not be your plan. If you already have a nano-banana key from another provider, you'll be editing base_url anyway - this node's value is the shape, not the default endpoint.
• "Auto (From Image)" needs a reference. Pick it with no image wired in and the behavior falls back to... whatever the provider decides. Wire a reference in if you're relying on it.
• Credential hygiene, again. This is a phone-home-with-your-key node by design. Scoped key, don't paste into shared workflows.

Reach for this one over the TutuPort variant when the VectorEngine-style endpoint and the auto-aspect behavior fit your setup. If you just want the model and don't care about the provider, either wrapper gets you there - this one just has the smarter aspect handling.