🖼️ LLMs_Vison_GLM4

If a workflow is asking for 🖼️ LLMs_Vison_GLM4, you're looking at a fossil. This was the ComfyUI-LLMs pack's dedicated Zhipu GLM-4V vision node, and the 1.0.0 refactor deleted it along with the Ali and Gemini siblings. GLM-4V support is alive and well - it just lives inside the pack's single LLMs Vision Unified node now, behind a model_type dropdown.

What it was

LLMs_Vison_GLM4 was the pack's GLM-4V node, and its inputs show how different the old design was. Instead of an image tensor, it took an image_url string - you pasted a URL to an image on the web and the node fetched it. The default value was a sample PNG from mihoyo's website, which tells you exactly how much thought went into that placeholder. The other inputs were prompt, model_name (glm-4v or glm-4), and an api_key field on the node. Output was a STRING called GETPrompt.

Mechanically it called Zhipu's API - the GLM-4V family is Zhipu/Z.ai's vision-language line, from the same Tsinghua spin-off that later made a name for itself in open video with CogVideoX. The image was base64-encoded and sent to client.chat.completions.create on the Zhipu SDK.

The key-on-the-node detail is worth flagging, because it's the same trap the old Gemini node had: a key in a widget becomes part of the workflow, and workflows travel (embedded in PNG metadata, shared everywhere). The modern pack moved keys to settings.yaml, where they belong.

The replacement

Install the pack, then use LLMs Vision Unified:

• model_type = glm4,
• model = glm-4v (or whatever current GLM vision model you configure),
• wire an actual image tensor in - the URL-pasting era is over, which is honestly an improvement.

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

Put your Zhipu key (open.bigmodel.cn) under vision_models → glm4 in settings.yaml, restart ComfyUI, and the glm4 option appears in the unified node. Use the clone URL above, not the one in the README - the README misspells it as leoleexh. ComfyUI Manager ("ComfyUI-LLMs") sidesteps that entirely.

Gotchas

• Placeholder key. The sample ships xxxxxx for the glm4 block.
• The mihoyo default. If you're porting an old workflow, check you didn't inherit that hardcoded sample URL as a prompt - it's a dead artifact of the old node, not anything you want in your graph.
• Proxy. BigModel (Zhipu) is a Chinese API; outside China it can be flaky without a proxy, same as the DashScope story.
• It's a billed API. GLM-4V calls are metered. The unified node is convenient, but every run still costs you per image.