GLM扩写高质量提示词

The most interesting thing about this node isn't the node - it's the model it defaults to. glm-4v-flash is Zhipu's (智谱AI) free vision model, and this node wraps it as an image-to-prompt expander: feed it an image, get back a polished English prompt. If you're tired of paying per call for prompt expansion, the free GLM flash tier is a legit reason to reach for this one. It's the "GLM扩写高质量提示词" node, part of ComfyUI_LLM_Banana, and it's the sibling of the Comfly expander - same job, different provider.

One catch that's easy to miss: this node does not take an IMAGE wire. Its required input is image_base64, a STRING containing a base64-encoded image. The image input is optional and may not be wired the way you expect - check the brief carefully. You'll often want to base64-encode your image upstream, or pass an image_url instead. That's the kind of quirk that makes people bounce off this node on their first attempt.

How it works

It posts the image (base64 or URL) to Zhipu's OpenAI-compatible API (open.bigmodel.cn), with a system prompt chosen by the image_prompt_preset dropdown - the same six presets as the Comfly expander: generic English description, Flux, video, Flux Kontext, poster design, meme/sticker design. The model writes the comma-separated prompt, quality words appended, and hands it back as a string.

The inputs that matter

• model_name - defaults to glm-4v-flash (free). 21 options in the dropdown: glm-4v-plus, glm-4.5v, glm-4, glm-4-air, the dated flash variants, and more. Only the vision ones (glm-4v*) can describe an image.
• api_key - your Zhipu key from bigmodel.cn. It also reads ZHIPUAI_API_KEY from the environment or Glm_Config.json.
• image_prompt_preset - shapes the output dialect for your target model.
• image_base64 - the actual image, as base64 text. This is the required path.
• image_url (optional) - skip base64 and hand it a URL instead.
• user_prompt / user_requirement - content to include, and constraints to obey.
• temperature, seed - sampling knobs.

Output is expanded_prompt (STRING). Feed it into your text encoder.

Install

ComfyUI Manager, search ComfyUI_LLM_Banana, or:

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/ComfyUI_LLM_Banana
cd ComfyUI_LLM_Banana
pip install -r requirements.txt

The pack's requirements include zhipuai, so the SDK comes along automatically. Restart, find the node under Ken-Chen.

Gotchas

The base64 requirement is the number one stumble - if the node produces nothing, check you're actually feeding base64 text, not an image tensor. Free flash tier has rate limits, so batch your expansion carefully if you're processing a dataset. And like every node in this pack, a real API key is involved: don't share workflows with your key embedded.