The prompt: how models read text
Writing a prompt a model understands: tags or sentences, order, length, language.
In plain words
A model does not understand text the way a person does. It was trained on pictures with captions and learned which words go with which pictures. Your prompt pulls the result towards pictures whose captions looked like it. So:
- a word that was rare in the captions does little, however exact it sounds;
- a prompt is not an order: every word pulls its own way, and strong words outpull weak ones;
- what you leave unsaid, the model decides for itself.
Two ways of reading
How you write depends on the text encoder of your model.
- Tags and short phrases (CLIP readers): SD 1.5, SDXL and the lines built on it. Example:
portrait of a woman, red coat, night street, rain, soft light. Anime lines such as Pony and Illustrious were trained on booru-style tags (1girl, red coat, outdoors, night, rain), and Pony also expects its own quality tags at the start (score_9, score_8_up, ...); the model page lists them. - Whole sentences (T5 and language-model readers): Flux, SD3, Qwen-Image, Z-Image, Wan, HunyuanVideo, LTX. Example:
A woman in a red coat stands on a night street in the rain. Soft light from a shop window falls on her face.These models follow who does what and where, and Qwen-Image and Flux can even write short text into the picture if you put it in quotes.
Order and length
- Put the main thing first: the subject, then the action, then the place, the light, the style.
- CLIP reads in chunks of 75 tokens (about 50 English words). Words in a later chunk hold together with the start less well.
- Sentence readers take a few hundred tokens, yet they too hold the end more weakly.
- Ten exact words work better than forty random ones. Every extra word takes attention from the rest.
- Most models were trained on English captions. Wan and Qwen-Image also understand Chinese well. Other languages work noticeably worse.
For video
A clip also needs what changes: one action, the camera's move, the pace. See the lesson A prompt for video and Video models.
Typical mistakes
- Negations in the main prompt:
no hattends to add a hat. Put unwanted things in the negative prompt. - Syntax from other programs:
<lora:name:1>,BREAK,[word]do nothing in ComfyUI. See Prompt syntax and weights. - A tag list for a sentence reader, or a novel for CLIP.
- Colours and details bleed between objects in tag models: describe each object in its own phrase.
How the app helps
The Check panel reads prompts the way ComfyUI will and lists "Prompts: worth changing": syntax from other programs, unmatched parentheses, weights above 1.5, the same word in both prompts, negations, non-English text, prompts too long for the encoder, a missing LoRA trigger word. The Models tab card shows the trigger words and example prompts the author stored in the file.