Plain-English explanation
A prompt often contains competing requests. Prompt weighting lets a user say that one phrase should matter more than another without deleting either. In a diffusion interface, emphasis may be expressed with parentheses or an explicit value, while de-emphasis may use brackets or a value below the neutral weight.
The control modifies conditioning rather than editing pixels directly. It cannot guarantee that the desired feature appears, and an overloaded prompt can remain internally contradictory even after weights are changed.
How it affects conditioning
Text is converted into embeddings by a text encoder such as CLIP. A weighting implementation scales or otherwise combines the representations associated with selected text spans before or during the model's denoising process. Exact handling differs: some interfaces parse nested punctuation, some support explicit numeric weights, and some pipelines offer longer or composable weighted-prompt encodings.
Because tokenization splits text into pieces, weighting a visible word is not always equivalent to changing one indivisible concept. Models also learn associations between concepts, so increasing a phrase can unexpectedly affect style, composition, or nearby attributes.
Where users encounter it
AUTOMATIC1111 documents attention syntax for increasing and decreasing emphasis, and comparable controls appear in ComfyUI workflows and hosted generators. Users may weight a character attribute, visual style, camera term, or element that the model otherwise ignores. It is most useful as a small correction after clarifying the prompt, not as a substitute for compatible models or reference conditioning.
Prompt weighting versus guidance scale
Prompt weighting changes the balance among pieces of conditioning. Guidance scale changes the overall strength of conditional guidance relative to an unconditioned prediction. A negative prompt supplies undesired conditioning; it is not simply a positive phrase with a smaller weight.
Limitations
- Syntax is not universal and can be treated as literal punctuation by an incompatible system.
- Extreme weights can create artifacts, oversaturation, or less natural compositions.
- A model cannot reliably express a concept it has not learned merely because that text receives a large number.
- Results also depend on seed, model, sampler, scheduler, and other conditioning inputs.