AI多模态助手

This is the flagship node of the ComfyUI-My-Nodes pack, and the one that shows you the whole point of the thing. It's a single node that feeds text, images, audio, or a local video file to Alibaba's qwen-omni-turbo model and gets back a written answer - or, if you flip one switch, a spoken one. No model download, no VRAM, no local inference at all. Everything is handled by an API call to Alibaba's Model Studio (DashScope), so this is squarely in the "API-wrapper node" camp: it looks like any other node on the canvas, but underneath it's an HTTP client with your key baked in.

The name is honest about what makes it worth installing: multimodal. Where most LLM nodes only take a text box, this one takes an IMAGE tensor, an AUDIO tensor, or a video path string and lets the model actually look at it. The mode dropdown is the heart of it - pick AI翻译, AI翻译+润色, 主题创意 (creative ideas), or the three "understanding" modes: 图片理解, 视频理解, 音频理解. In those three, an empty text box defaults to prompt reversal - the model describes what it sees and hands you back a ready-to-use prompt. That's the same captioning job the KB tracks as the community's bread and butter (see llm-in-comfyui.md on VLM captioning), just routed through an API instead of a local Qwen-VL.

The mechanism is straightforward under the hood: an image tensor gets converted to a PIL image, saved to a temp PNG, base64-encoded, and shipped to DashScope's OpenAI-compatible endpoint. For audio replies, the streamed response is decoded into a WAV and loaded back as an AUDIO tensor. The outputs are OutText (STRING) and OutAudio (AUDIO) - so you can wire the text straight into a prompt node or the audio into a preview.

The few inputs you actually set: api_key (see install below), mode, out_language (English or Chinese), and out_audio. Be warned - the author's own tooltip on out_audio says turning it on "makes output possibly uncontrollable," which is code for "the model decides what it says, and sometimes it rambles." audio_voice picks the voice (Cherry, Serena, Ethan, Chelsie).

Install - from ComfyUI Manager, search "ComfyUI-My-Nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Tagbliton/ComfyUI-My-Nodes

then restart ComfyUI. It needs an API key; the recommended path is dropping it in the pack's config.json and reading it with the "从配置文件获取数据" node, or setting the DASHSCOPE_API_KEY environment variable.

Where people get burned: the README's known-issue list is worth reading before you blame yourself. Connection error means network/proxy trouble, not a bad key. And the 'NoneType' object has no attribute 'is_cuda' error in image-understanding mode is a known bug - the README's own fix is to use the dedicated "AI图片理解" node instead. The bigger gotcha is upstream: DashScope keys need a real-name-verified mainland China account, so people outside China hit Invalid Api-Key provided no matter how carefully they paste. That's an Alibaba account problem, not a node problem - check that before you debug anything else.