Prompt syntax and weights
The few special forms ComfyUI really understands in a prompt, and the forms from other programs that it does not.
What ComfyUI understands
| Form | What it does |
|---|---|
(red coat:1.2) | Weight: the words in the parentheses pull with 1.2 instead of 1. Below 1 weakens: (rain:0.8). |
(red coat) | Weight 1.1 without a number. Nested parentheses multiply: ((red coat)) is about 1.21. |
\( and \) | A plain parenthesis that must stay in the text, for example in a tag like artist \(style\). |
embedding:name | Uses a textual inversion file named name from models/embeddings. |
{day|night|dawn} | A dynamic prompt: the ComfyUI editor picks one option each time you queue. \{ keeps a plain brace. |
// ... and /* ... */ | Comments: the editor removes them before the prompt is encoded. |
Weights, embeddings and escaped parentheses are read by ComfyUI's text encoding. Curly braces and comments are handled by the ComfyUI editor when the prompt is queued, so a script that sends the API format straight to the server does not get them.
What it does not understand
These come from other programs (mostly the AUTOMATIC1111 and Forge web UIs) and are plain characters in ComfyUI:
<lora:name:0.8>: connects nothing. A LoRA is added with a LoRA loader node.BREAK,AND: not special.[word](weaken in other programs),[a:b:10](prompt editing): not special. Prompts copied from picture sites are often full of such forms. Take them out.
Values and what they change
- Keep weights between about 0.5 and 1.5; 1.1 to 1.3 is usually enough. Above 1.5 pictures burn and distort.
- ComfyUI applies weights more strongly than some other programs, so a prompt copied with its weights often needs them lowered.
- On sentence readers (Flux, Qwen-Image, Wan and others) weights act weakly and unpredictably. Say what matters in words and put it first.
Typical mistakes
- Unmatched parentheses: the weight lands on other words than you meant.
- A weight of 2 or 3 copied from elsewhere: a burnt picture.
embedding:with a file of another family: no effect or a warning.
How the app helps
Check finds LoRA tags, BREAK, AND and square brackets, unmatched parentheses, weights above 1.5, and weights used on a model that reads sentences. The prompt lesson has a step where you set a weight yourself.