💬GLM4_Vsion_IMGURL

This is the interesting node in the pack. GLM4_Vsion_IMGURL takes an image URL plus your instruction, sends both to Zhipu's glm-4v vision model, and returns the model's text. The author's intended use is right there in the default prompt: caption a reference image into an English, comma-separated prompt for your diffusion model - "best quality, high resolution, 4k" tags and all. It's a quick-and-dirty LLM captioner inside ComfyUI, no local model required.

How it works

The node builds one user message with two parts - a text block (your prompt) and an image_url pointing at the image - and calls chat.completions.create against glm-4v. The model reads both, and its reply comes out the GETPrompt output as a string. Clean and simple, as long as you respect the one big constraint below.

The gotcha that burns everyone

The image must be at a public URL. The README is explicit: base64 upload to GLM-4 was too slow, so "we temporarily support uploading via URL." "Temporarily" was written in January 2024, and the pack hasn't been touched since February 2024 - so it's still URL-only, and probably always will be.

What that means in practice: this node can't see your local files, and it can't see what's already sitting in your ComfyUI graph. You have to host the image somewhere public - imgur, a GitHub raw link, any static host - and paste the URL. A file:// path or a locally loaded image won't work. That's the thing that trips up everyone on their first try, because the node doesn't look like a web tool; it just is one.

The inputs that matter

Four required fields:

• prompt (multiline STRING) - your instruction. The shipped default is Chinese and reads like an SDXL-captioning spec: describe as many details as possible, write English comma-phrases, add [[best quality, high resolution, 4k, high quality]], third person. You can replace it with anything - "write a negative prompt for this image", "describe the lighting and composition", whatever the job needs.
• image_url (STRING) - the public URL of the image. This is the field that defines the node.
• model_name - locked dropdown, glm-4v only.
• api_key - your Zhipu key, defaulting to the pack's config.json placeholder.

Output: GETPrompt (STRING) - the model's caption or answer. Feed it into your positive prompt, or into a text input that leads toward the encoder.

Two honest caveats

The default prompt is tuned for the SDXL/anime tag dialect, where the best quality boilerplate is still alive and doing something. If you're targeting a newer LLM-encoded model, that tag list is mostly inert - rewrite the prompt for what you're actually feeding.

And every run is one API call that sends your image off your machine. If you caption a handful of images, fine. If you caption hundreds, the cost adds up and your images are leaving the building. The local GLM-4 wrapper route (running glm-4v-9b on your own GPU) is the free and private alternative; this node is the wire-it-in-now choice.

Install

The whole pack installs at once - all three nodes ship in one repo:

cd ComfyUI/custom_nodes
git clone https://github.com/JcandZero/ComfyUI_GLM4Node
cd ComfyUI_GLM4Node
pip install -r requirements.txt   # just installs `zhipuai`

Or via ComfyUI Manager: search "GLM4Node", install, restart. The node lands in the BlinkNodes_PROMPT category. No model downloads, no VRAM - just a Zhipu API key, internet, and a publicly reachable image. And remember the README's warning: the key in the author's demo video is dead, so get your own.