扩写高质量提示词 (Comfly/T8)

You've got a reference image and you want a proper prompt for Flux, LTX, or whatever base you're running - the kind of comma-separated, detail-loaded prompt that actually steers a model instead of vague hand-waving. This node is that shortcut: give it an image (or an image URL), it sends the image to a vision LLM on a Comfly/T8 mirror API, and out comes a polished English prompt. "扩写高质量提示词" is the display name, which translates roughly to "expand high-quality prompt," and that's exactly the job.

It belongs to the Comfly/T8 family inside ComfyUI_LLM_Banana, so it's an API-wrapper node rather than a local one. No weights to download, no VRAM - but the image and prompt go over the wire to a third-party service, and the call costs you whatever your mirror charges.

How it works

Same machinery as the sibling Comfly_First_Last_Frame_Prompt: the image gets base64-encoded and POSTed to {base_url}/chat/completions with a Bearer key. What's different is the prompt construction - the node comes with an image_prompt_preset dropdown that swaps in a tailored system prompt, so the output is shaped for your target rather than generic.

The inputs that matter

• image_prompt_preset - the six presets are the real feature: 通用高质量英文描述 (generic high-quality English description), Flux描述 (Flux), 视频描述 (video), Flux Kontext, 海报设计 (poster design), 表情包设计绘图 (meme/sticker design). Pick the one that matches where the prompt is headed. The Flux and Flux Kontext presets matter because those models want their own prompt dialect.
• image - optional IMAGE wire, if you're pulling from a Load Image node.
• image_url / ref_image - alternatives to the IMAGE wire, for when your source is a URL or a text path.
• user_prompt and user_requirement - one adds content ("there's a red car in the foreground"), the other adds style constraints.
• api_provider, api_key, base_url - the Comfly/T8 mirror plumbing, same as its sibling node.

Output is a single expanded_prompt STRING. That's your text-to-image prompt, so wire it into the CLIP/text-encode stage of your generation workflow.

Install

Via 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

Then restart. The dependency list is heavy - transformers, opencv-python, google-genai, openai, httpx and friends.

Gotchas

If the node spits out an error string in place of a prompt ("未检测到API Key或Base URL…"), your key/base URL aren't resolving - configure ChatFly_config.json under prompt_api_providers, or fill the node's fields. And the usual caveat for this whole pack: it's a fresh-ish Chinese-community pack with near-zero English reddit footprint, so treat it as a convenience tool rather than a vetted institution, and don't leave a real API key lying around in shared workflows.