ModelScope-Text 文本生成节点

Somewhere between "prompt here" and "image there," almost every ambitious ComfyUI workflow wants a brain: something that rewrites a vague idea into a strong prompt, expands a sentence into a scene, or translates a tag soup into clean natural language. Locally that means running a 7B model just for text duty. This node skips that - it's a chat-completion client for ModelScope's hosted inference, so you get DeepSeek-V3.1, Qwen3-Coder-480B, GLM-4.5, or step3 without a single local model file.

How it works

It works the way every OpenAI-SDK call works: you hand it a user_prompt and an optional system_prompt, it POSTs to https://api-inference.modelscope.cn/v1 as an OpenAI-compatible chat request, and returns the assistant's text as a STRING output named response. That string feeds directly into any prompt input in your graph - chain it into the ModelScope-Image node and you've built a text-to-prompt-to-image pipeline where the LLM does the prompting.

The inputs that matter

Inputs to care about:

• user_prompt and api_token (both required) - the token comes from modelscope.cn and auto-saves to modelscope_config.json on first use, so later runs can leave it blank.
• system_prompt - defaults to "You are a helpful assistant." This is the field that turns the node into a specialized tool. Set it to "You are a prompt engineer. Rewrite the user's idea as a detailed Stable Diffusion prompt." and the node becomes a prompt booster.
• model - a 17-entry dropdown. Default is Qwen/Qwen3-Coder-480B-A35B-Instruct, which is a code model, a slightly odd default for prompt work - switch it to deepseek-ai/DeepSeek-V3.1 or a Qwen3-Instruct variant unless you actually want code. The list comes from the config file, so it's editable.
• max_tokens (100–8000, default 2000) and temperature (0.1–2.0, default 0.7) - standard. Crank temperature for creative prompt variation, keep it low for structured rewrites.
• stream (default true) - streams the response into the console as it arrives. Handy feedback during long generations; harmless to leave on.
• seed - here's the honest bit: this is largely placebo. The API doesn't accept a seed for chat completions; the code just seeds numpy's RNG before the call, which has no effect on the model's output. Treat it as a "make a different random-ish answer" button, not a reproducibility control.

The same "error as output" quirk from the rest of the pack applies: failures come back as the response text with "文本生成失败:" prefixed, and the real trace lands in the terminal. And the openai package is required - if it's missing, the node turns into a placeholder telling you to install it.

Installing it

Install the pack itself via ComfyUI Manager (search "ComfyUI ModelScope API Node") or:

cd ComfyUI/custom_nodes
git clone https://github.com/hujuying/ComfyUI-ModelScope-API

then restart ComfyUI.

The bottom line

Reach for this when you want prompt engineering inside the graph without babysitting a local LLM, or when you're on a machine that can't run a 7B comfortably alongside diffusion. Skip it if you already have a local LLM wired into your workflow - for pure prompt-rewriting, an 8B on your own GPU is faster than any round trip, and doesn't bill per call.