💎 Gemini 多功能 @炮老师的小课堂

The "多功能" (multifunction) in the name isn't marketing fluff - this node genuinely does four jobs. Gemini3_Multimodal ("💎 Gemini 多功能 @炮老师的小课堂") is a Google Gemini client that handles plain LLM chat, multi-image analysis (image captioning/反推), video analysis, and audio analysis, all from one socket that outputs a response string. It's the pack's "put a VLM in my graph" node, and because it's Gemini, everything is a cloud call - no local model, just your key.

How it's set up

The interesting bit is the config. The node reads gemini3_config.json (in the pack folder) to know which API providers exist, which models each exposes, and the default provider. Out of the box the providers are google, comfly, and T8 - the "镜像站" (mirror-site) dropdown on the node lets you switch between comfly / hk / us / 柏拉图 endpoints. In plain terms: it's built for the reality that Google's API isn't reachable or payable from everywhere, so you can route through a mirror or third-party provider. If you only have a standard Google API key, use the plain Google provider and don't touch the mirrors.

The inputs that matter:

• 🎯 系统角色 and 💬 用户输入 - the two text boxes. Chat-only mode needs nothing else.
• 🤖 模型选择 - a string, default gemini-3.1-flash-lite-preview.
• 🌐 镜像站 - comfly / hk / us / 柏拉图, matching whatever provider config you loaded.
• 🔑 API密钥 - the provider's key.
• 📊 输出语言 - 中文 or 英文, handy because the node's default persona is Chinese.

Optional media: 🖼️ 图像1–4, 🎬 视频 (note: typed IMAGE in the schema, so feed it accordingly), 🎵 音频 (needs scipy - the pack's requirements.txt installs it for exactly this), plus 🌡️ 温度, 🎲 top_p, 📝 最大令牌 for generation control.

Outputs

Just the one: response (STRING). It's the whole point - one socket that gives you back the model's answer to whatever multimodal job you set up. Wire it into a text display, or chain it into a prompt enhancer pipeline if you're using it for captioning.

Install and honest notes

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

Restart. The async machinery (aiohttp, nest-asyncio) is bundled in the pack's requirements, so there's no manual step beyond the key.

The practical framing: if you already run a dedicated Gemini wrapper pack, this is a Chinese-UI alternative with the mirror-site flexibility baked in - genuinely useful if you're in a region where the official endpoint is a problem. If you're just after occasional captioning, it's overkill and you're paying per call. One caveat worth repeating from the KB: this is an API node that holds a key and phones out by design, and it reads a config file to decide where "out" is - so if you edit gemini3_config.json yourself, make sure you know which endpoint you just pointed it at. The mirrors exist for region access; they don't change what Google's model will refuse to say.