🎨通用图像编辑API (测试版) @炮老师的小课堂

Some nodes are opinionated about the vendor. ImageEditAPINode ("🎨通用图像编辑API (测试版) @炮老师的小课堂") is opinionated about the workflow instead: it wraps a chat-completions-style multimodal image edit API - the kind Gemini 3 / Nano Banana 2 use - and layers on a pile of quality controls: aspect ratio, output resolution, quality presets, style presets, and an AI upscale stage with nine different upscale models. It's marked 测试版 (beta), and it shows: the default endpoint is a third-party relay (api.tu-zi.com), the model default is gemini-3-pro-image-preview, and the README points at a wiki page for that model. This is a "point it at whatever speaks this dialect" node with extra frosting.

The controls that matter

• 💬 提示词 - the edit instruction. Plus 🎯 系统角色 if the endpoint supports it.
• 🌐 API地址 / 🔑 API密钥 / 🤖 模型名称 - the trinity. Default model is gemini-3-pro-image-preview (Nano Banana 2 class), which is where the node's feature list makes most sense.
• 📡 请求方法 / 🔐 密钥位置 / 📝 密钥字段名 - POST default, key in Header, field Authorization. Same flexible auth pattern as the universal API node.
• 🖼️ 图像1–4 - up to four reference images for the edit.
• 📐 宽高比 - Auto plus ten ratios.
• 📊 响应模式 - TEXT_AND_IMAGE or IMAGE_ONLY.
• 🖥️ 输出分辨率 - Auto / 1K / 2K / 4K. Tooltip is explicit: only supported by Nano Banana 2 (Gemini 3) models. On other models, leave it at Auto or you'll get a rejected call.
• 🎨 画质预设 - standard / hd / ultra_hd / ai_enhanced / ai_ultra.
• 🎭 风格预设 - vivid / natural / artistic / cinematic / photographic.
• 🔍 放大倍数 - 1x (不放大) / 2x / 4x / 6x, with 🧩 放大模型 (High Fidelity, Standard, Art & CG, Lines, Very Compressed, Low Resolution, Text & Shapes, Redefine, Recover). The upscale runs as part of the same call.

Optional plumbing: 🎯 响应提取路径 (jq-style path into the response JSON), ⏱️ 超时时间, 📋 额外Headers, **📦 额外Body参数`. Outputs are the useful quartet: image (IMAGE), response (STRING), image_url, raw_json.

The honest picture

This node is a good example of the reseller ecosystem in action - it ships pointed at a relay, because that's how many people in this space reach Nano Banana 2 outside Google's own billing. That's fine, but it's exactly the scenario the KB's external-API essay flags: a node that holds your key and phones out by design. So treat the default endpoint as a suggestion, check what you're actually sending where, and remember the beta label means the mapping of presets to endpoints can change without notice. On the practical side: the 输出分辨率 field will bite you first - leave it on Auto unless you're certain the model supports 2K/4K. And if a call fails, raw_json is where the endpoint tells you what it didn't like.

Install

cd ComfyUI/custom_nodes/
git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI.git
cd ComfyUI-dapaoAPI
pip install -r requirements.txt

Restart ComfyUI. No models download - the "models" here live on the endpoint you configure.