🖼️ LLMs_Vison_Ali

If you landed here because a workflow is screaming that 🖼️ LLMs_Vison_Ali is missing - relax, nothing about your image pipeline is broken. This node was removed in the ComfyUI-LLMs pack's 1.0.0 refactor, which swept away the pack's separate per-provider vision nodes and replaced them with a single LLMs Vision Unified. Alibaba's vision models didn't go anywhere; they just moved behind one dropdown.

What it was

LLMs_Vison_Ali was the dedicated Alibaba Qwen-VL node. It took an image, a text prompt, and a model choice (qwen-vl-plus or qwen-vl-max), and sent them to Alibaba's DashScope API - the cloud service where the Qwen team (Alibaba's Tongyi Lab, the same outfit behind Wan and Z-Image) hosts its vision-language models. Notably it also had a seed input, a leftover from an older design that the refactor quietly dropped. Output was a STRING named text, i.e. the model's description of your image.

The mechanism is unchanged in spirit: image gets base64-encoded and shipped to MultiModalConversation.call() on DashScope, and the returned text comes back into your graph. No local weights, no VRAM - you were renting Alibaba's GPU the whole time.

The migration

Today you don't install a per-provider node at all. Install the pack, and use LLMs Vision Unified:

• set model_type to ali,
• set model to qwen-vl-plus (fast, cheap) or qwen-vl-max (smarter, pricier),
• wire your image in and write your prompt.

That node lists every provider's models in one dropdown, so make sure model and model_type actually match - pick qwen-vl-plus while model_type says gemini and it returns an error string instead of a caption.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/leoleelxh/ComfyUI-LLMs
cd ComfyUI-LLMs
pip install -r requirements.txt
cp settings.yaml.sample settings.yaml

Then in settings.yaml, under vision_models → ali, put your DashScope API key (get one at dashscope.aliyun.com). Run python validate_config.py and it'll confirm the ali provider is configured. The README's own clone URL is misspelled (leoleexh vs leoleelxh), so use the one above or ComfyUI Manager → "ComfyUI-LLMs".

Gotchas that survived the rename

• Placeholder key. The sample ships sk-xxxxx for the ali block; a real key is a must, and the DashScope key format changed over the years - if you have an old one, regenerate it.
• Proxy. DashScope is Alibaba Cloud, which outside mainland China sometimes means routing issues. The README's "some models may require proxy access" covers exactly this.
• It's a paid API. Every run is a billed Qwen-VL call. For batch captioning, that adds up fast - think before you blast a 64-image batch through it.