GLM文本对话

What it actually does

You type "a small dog playing on a grassy lawn" into a box, and this node hands you back a full paragraph - the dog's coat, the light, the camera angle, the mood - ready to paste into a CLIP Text Encode. That's the whole job, and it's the same prompt-expansion trick LLM-in-graph workflows have been doing for years. No model download, no VRAM, no GPU. GLM_Text_Chat is an API wrapper: it packages your text, phones home to Zhipu AI's GLM API (the Z.ai behind the open GLM and CogVideoX lines), and drops the reply back into your graph as a plain string.

Why you'd reach for it (and the free bit)

The usual advice for prompt enhancement is a local abliterated 8B model - offline, uncensored, free per call. But local means budgeting VRAM for a second model that has to share your card with the diffusion model. This node takes the opposite door. It's an API call, and here's the thing the README leads with: the default model, GLM-4.5-Flash, is free. Zhipu's Flash tier is the giveaway model line, and the whole pack is built around it. So you get a cloud LLM inside the graph that costs nothing per call and eats no VRAM. That's the reason this pack exists.

The honest caveats come straight from the API-node playbook: your prompt leaves the machine, and Zhipu's model runs its own content filter. Don't wire anything through it you wouldn't paste into a cloud chatbot.

The inputs that matter

• text_system_prompt_preset - the instruction the model follows. Three to pick from: 视频扩写 (video prompt expansion), Flux扩写 (FLUX-style expansion), and Flux Kontext (command-style prompts for FLUX Kontext's dual-image editing). Default is video expansion.
• system_prompt_override - fill this in and it beats the preset. That's the escape hatch when the built-ins aren't your style.
• api_key - leave it blank and the node falls back to the ZHIPUAI_API_KEY env var, then to config.json in the pack folder.
• model_name - defaults to GLM-4.5-Flash. Keep it on the Flash line or the "free" part of this story dies.
• text_input - your rough idea, multiline.

There's also temperature (0.9), top_p (0.7), max_tokens (1024) and seed. Here's the trap: the seed is cosmetic. The tooltip says it flat out - it only affects randomness inside the node, not the Zhipu model's output, because the API doesn't accept a seed. Set it to anything; nothing reproducible happens.

The single output is response_text, a plain STRING. Wire it into a CLIP Text Encode, a text display node, or any STRING input. The common move is chaining it straight into a FLUX or video sampler.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/msola-ht/ComfyUI-GLM4.git
cd ComfyUI-GLM4
pip install -r requirements.txt

Then restart ComfyUI. ComfyUI Manager also has it - search "ComfyUI-GLM4". Dependencies are light: zhipuai, Pillow, numpy - all pip-installable, no torch or CUDA wrangling (ComfyUI already ships those).

Then the key: sign up at bigmodel.cn, generate an API key, and paste it into the node, set ZHIPUAI_API_KEY in your environment, or drop it into the pack's config.json.

Troubleshooting

• "API Key 未提供" comes out as the node's output text. The key didn't resolve anywhere. Check the node field, env var, and config.json. Note the pattern: this pack returns errors as its output string rather than crashing ComfyUI. Easy to miss, but genuinely friendlier than a hard crash.
• The free tier stopped being free? Only if you wandered off the Flash models. The README bumped GLM-4.5-Flash to default; stick to the Flash line and you're on the free quota.
• Chinese display names. The node shows up as "GLM文本对话" and the README is Chinese. English speakers end up navigating by class name (GLM_Text_Chat), because that's what the search and save function use.
• Maintenance is thin. This is a one-person hobby pack. The README's own changelog admits Zhipu updated their Python library and the plugin "still works but needs a rebuild." If a future API change breaks it, don't be shocked.

Set expectations while you're at it: it's a prompt expander, not a writer. Competent, structured detail - not GPT-4-class prose. For zero VRAM and zero dollars, that's a fair trade.