Prompt Weighting Syntax: (word:1.2), [word] and :: Explained

Prompt weighting tells a model to pay more or less attention to part of a prompt. The catch is that every tool has its own syntax, and in some tools brackets do nothing at all. This page gives the syntax per tool, then practical ranges.

Cheat sheet

AUTOMATIC1111

The AUTOMATIC1111 wiki documents the attention syntax:

a (red) car            -> red × 1.1
a ((red)) car          -> red × 1.21 (1.1 × 1.1)
a [red] car            -> red ÷ 1.1
a (red:1.5) car        -> red × 1.5
a (red:0.25) car       -> red × 0.25 (four times weaker)
a \(red\) car          -> literal brackets, no weighting

Numeric weights only work inside round brackets. Square brackets cannot take a weight: [word:0.5] is prompt editing (it adds "word" after half the steps), not a down-weight.

Practical examples:

portrait of a woman in a (red:1.3) coat, snowy street, soft light

a cozy cabin interior, (fireplace:1.2), (window:0.8), warm evening light

(masterpiece, best quality:1.1), a lighthouse on a cliff at sunset, (dramatic clouds:1.25)

Related syntax in the same tool

These are not weights but often get confused with them:

ComfyUI

ComfyUI's README documents the same round-bracket syntax in the text encode node: (good code:1.2) or (bad code:0.8), and plain (word) defaults to ×1.1. Two differences from A1111 that trip people up:

• [word] is not a documented down-weight in ComfyUI. Use (word:0.9) instead.
• {day|night} is a random choice each time you queue the prompt, not a weight. Escape literal braces as \{ \}.

a (golden retriever:1.2) running on a beach, (wet sand:0.9), sunrise, cinematic

Diffusers with Compel

Plain Diffusers pipelines do not parse (word:1.2); the brackets go to the text encoder as ordinary characters. The Compel library adds a weighting syntax and returns prompt embeddings that you pass to the pipeline:

 compel  CompelForSD

compel = CompelForSD(pipeline)
conditioning = compel()
image = pipeline(prompt_embeds=conditioning.embeds).images[]

+ raises the weight of the word before it, - lowers it, and repeating the symbol makes the effect stronger. Compel's documentation also covers SDXL and FLUX pipelines, which need pooled embeddings.

Midjourney

Midjourney uses multi-prompts: a double colon :: splits the prompt into separate ideas, and a number right after it sets how important that part is. Midjourney's own example:

space::2 ship

Here "space" counts twice as much as "ship", so the image leans toward space scenery rather than a spaceship. Parts without a number have a weight of 1. Midjourney's --no parameter documentation says that --no plants is the same as giving "plants" a weight of -0.5.

Check the compatibility note on Midjourney's multi-prompt page before relying on ::: the help center marks the feature as not supported in some newer model versions. In those versions, use clear wording and --no instead.

GPT Image, FLUX and other natural-language models

There is no documented weight syntax for OpenAI's GPT Image models or for FLUX. They read prompts as sentences, so emphasis comes from wording:

The most important thing in this image is the bright red umbrella. A woman holds it on a grey, rainy street; everything else is muted.

• Put the key element first.
• Say it explicitly: "the main focus is…", "clearly visible…".
• State what should be absent as a sentence: "no text anywhere in the image".

Safe weight ranges

These are rules of thumb, not limits. How far you can push depends on the model, the sampler and the CFG scale.

Common mistakes

• Using A1111 syntax in the wrong tool. In Midjourney or a plain Diffusers script, (word:1.3) is just text.
• Stacking brackets. (((((word))))) is ×1.61 and hard to read. Write (word:1.6) if you really mean it, but usually you don't.
• Weighting everything. If every term has a weight, nothing stands out. Weight one or two elements.
• Fixing with weight what belongs in the negative prompt. To remove something, use the negative field. See the negative prompt lists.
• Forgetting regional control. If two subjects bleed into each other, weighting won't separate them. Try regional prompting.

For SDXL prompt structure and settings, see the SDXL prompt guide. For random variation instead of weights, see the wildcards guide.

Build your prompt

The prompt generator builds a prompt from your subject, style and lighting choices that you can then weight in your own tool. It runs in your browser and does not generate images.

Primary sources

• AUTOMATIC1111 wiki: Features (attention/emphasis, prompt editing, BREAK, AND)
• ComfyUI README: prompt weighting and dynamic prompts
• Compel: prompt weighting for Diffusers
• Midjourney: Multi-Prompts & Weights
• Midjourney: No parameter