A language model writes the prompt

<!-- lesson: llm-prompt -->

Lesson "llm-prompt", in the group Packs and tools of the lesson list. Steps: 10.

#What you will be able to do

You let a language model inside ComfyUI turn a short idea into a detailed prompt, and compare it with your own by the switch.

#What you need

• It goes on from Two versions of a prompt with one switch. The card says so; it is a hint, not a lock.
• Model files the lesson checks:
  • for clip_name of Load CLIP: a file whose name has qwen or gemma.
• Nodes: CLIP Text Encode (Prompt), If/Else Switch, Text (Multiline), Load CLIP, Generate Text / TextGenerate, Preview as Text.
• The app downloads and installs nothing. The step texts below say which files and folders, and which ComfyUI template names them.

#The steps

The words of each step, as the lesson shows them. Node names are in bold.

1. A language model in the graph (reading)

   A language model can turn a short idea into a detailed prompt: light, composition, materials, style. ComfyUI does it with its own Generate Text node, on your graphics card, without another program and without a cloud service.

   This lesson goes on from "Two versions of a prompt with one switch": your idea stays on one side of the switch, and the model's version goes to the other.

   It needs a language model file in ComfyUI/models/text_encoders: one that ComfyUI's text generation templates name, such as Qwen 3 4B, Qwen 3.5 or Gemma 4. Take it from the model's official page; the app downloads nothing.

2. The switch graph is needed

   The canvas must have the graph of "Two versions of a prompt with one switch": If/Else Switch with its output in CLIP Text Encode (Prompt), and the Text (Multiline) node with your prompt on its on_false input.

   If it is not there, do that lesson first.

3. Load the language model

   Add a Load CLIP node and choose the language model file in its clip_name field. Leave type as the template of that model sets it; the official Qwen 3.5 and Gemma 4 templates leave stable_diffusion.

   A language model is loaded like a text encoder: in ComfyUI the two are the same kind of file. This node is not linked to the text encoder of the picture model: it feeds only Generate Text.

4. Generate Text

   Add a Generate Text node. Connect the output of Load CLIP to its clip input and the output of Text (Multiline) to its prompt input. Set max_length to 512.

   The model now reads your idea. max_length is the longest answer in tokens, pieces of words; a prompt needs far fewer, the limit only stops a model that talks on. Leave thinking off: the answer is then the prompt itself.

5. The model's version into the switch

   Connect the generated_text output of Generate Text to the on_true input of If/Else Switch. The link from the second version is replaced.

   Now true sends the model's prompt to the picture and false your own. The second text node is free: the next step gives it a new job.

6. The instruction

   Connect a Text (Multiline) node (the free one, or a new one) to the system_prompt input of Generate Text and write the instruction into its value field. For example: Rewrite the idea as one detailed image prompt in English: subject, setting, light, camera, style. Answer with the prompt only.

   The instruction says what to do with the idea; the idea says what about. "Answer with the prompt only" keeps greetings and explanations out of the picture.

7. See what the model wrote

   Add a Preview as Text node and connect the generated_text output of Generate Text to its source input.

   After a run the text shows on the node. Read it: if the model adds things you did not want, say so in the instruction.

8. Memory and time (reading)

   The language model is a model like any other: a 4B file takes about 8 GB in 16 bits, less in fp8. ComfyUI loads it, writes the text, and moves it away when the picture model needs the room.

   With a fixed seed and an unchanged idea ComfyUI keeps the written text and does not run the language model again. Change the idea, and only then is a new text written.

9. Other ways (reading)

   • comfyui-ollama: sends the idea to an Ollama server you run yourself; Ollama Generate gives the text from its result output.
   • ComfyUI-QwenVL: Qwen vision-language models that describe a picture; the pack downloads them into models/LLM on the first run.
   • comfyui_LLM_party: agents and cloud models with your API key.
   • Partner nodes of ComfyUI (OpenAI, Gemini and others) run in the cloud.

   Cloud models cost money and send your text to that company. Generate Text runs on your own card.

10. The graph is ready (reading)

Save the graph with "File ▸ Save workflow for ComfyUI…", open the file in ComfyUI and run it there. This app runs nothing in ComfyUI itself.

Run once with true and once with false on the switch, with the same seed. Keep what works: the short idea, the long prompt, or the model's prompt edited by hand.

#What the lesson checks

A reading step is done when you press Got it, next. A doing step ticks itself off when the canvas holds:

• Step 2, The switch graph is needed:
  • a CLIP Text Encode (Prompt) node.
  • a If/Else Switch node.
  • a Text (Multiline) node.
  • output of If/Else Switch into text of CLIP Text Encode (Prompt).
  • STRING of Text (Multiline) (idea) into on_false of If/Else Switch.
• Step 3, Load the language model:
  • a Load CLIP node.
  • clip_name of Load CLIP holds a file whose name has qwen or gemma (capitals, dots and dashes do not matter).
• Step 4, Generate Text:
  • a Generate Text or TextGenerate node.
  • CLIP of Load CLIP into clip of Generate Text or TextGenerate.
  • STRING of Text (Multiline) (idea) into prompt of Generate Text or TextGenerate.
  • max_length of Generate Text or TextGenerate is from 256 to 1024; Do this for me sets 512.
• Step 5, The model's version into the switch:
  • generated_text of Generate Text or TextGenerate into on_true of If/Else Switch.
• Step 6, The instruction:
  • a Text (Multiline) node.
  • STRING of Text (Multiline) (instruction) into system_prompt of Generate Text or TextGenerate.
  • value of Text (Multiline) (instruction) holds at least 30 characters.
• Step 7, See what the model wrote:
  • a Preview as Text node.
  • generated_text of Generate Text or TextGenerate into source of Preview as Text.

#See also

• The prompt: how models read text
• Text encoders: CLIP, T5 and others
• How lessons work