Stable Diffusion prompt weights: how (word:1.2) works, and how far to push it

Prompt weights let you tell Stable Diffusion that one part of a prompt matters more, or less, than the rest. Used gently, they fix the common problem of a detail being ignored. Used heavily, they break images in very recognisable ways. This guide covers the syntax used by the popular interfaces, the ranges that behave well, and the habits that make weights mostly unnecessary.

The basic syntax

The most widely used syntax comes from the AUTOMATIC1111 web UI and is supported, with small differences, by Forge, ComfyUI and many other front ends.

• (word:1.2) sets an explicit weight. 1.0 is neutral, above 1.0 adds emphasis, below 1.0 reduces it.
• (word) is shorthand for a weight of 1.1.
• ((word)) multiplies twice, giving about 1.21. Each extra pair multiplies by 1.1 again.
• [word] reduces emphasis to roughly 0.91 in AUTOMATIC1111-style interfaces.
• (red scarf:1.3) weights the whole phrase, not just one word.

Explicit numbers are easier to read and adjust than stacked brackets. (lantern:1.21) and ((lantern)) do the same thing, but only one of them tells you the value at a glance.

If you need literal parentheses in your prompt, escape them with a backslash: \( and \). Otherwise the interface will treat them as weighting.

What a weight actually does

Your prompt is turned into a sequence of token embeddings by the text encoder. A weight scales the embeddings for the tokens inside the brackets, which increases or decreases their influence on the image. It doesn’t add information; it changes how loudly existing information speaks.

Implementations differ in the detail. AUTOMATIC1111 rescales the result so the overall strength of the prompt stays roughly constant, while ComfyUI’s default handling doesn’t, so the same number can feel stronger in ComfyUI. If you move a prompt between interfaces, expect to retune the weights.

Sensible ranges

There are no official limits, but in practice:

• 0.5 to 0.9: gently de-emphasise something that keeps dominating, such as a background element or an overpowering style word.
• 1.1 to 1.3: the useful range for most emphasis. Enough to rescue an ignored detail without side effects.
• 1.3 to 1.5: strong. Use it for one key element, and check the result closely.
• Above 1.5: usually causes more problems than it solves.

If something still doesn’t appear at 1.4, the weight is rarely the real problem. The concept may be weak in the model, conflict with another part of the prompt, or be buried at the end of a long prompt.

Over-weighting artefacts

Over-weighting has a recognisable look:

• Concept bleed: (red:1.8) turns the whole image red, not just the scarf.
• Repetition: the weighted object appears several times, or as a pattern across surfaces.
• Burned colours and contrast: saturated, crunchy patches similar to a CFG scale set too high.
• Broken anatomy and composition: the model sacrifices structure to satisfy the heavy term.

When you see these, lower the weight before you change anything else. Several moderate weights also add up: five terms at 1.4 can do as much damage as one at 2.0.

Order matters as much as weight

Stable Diffusion pays more attention to the start of the prompt. Moving a detail closer to the front often does what people reach for weights to do, without any artefacts.

Before: a cosy reading nook, armchair, bookshelves, warm lamp light, rain on the window, (ginger cat:1.4)
After:  a ginger cat curled up in an armchair in a cosy reading nook, bookshelves, warm lamp light, rain on the window

The second version has no weights at all, but the cat is now the subject of the sentence and comes first. Try reordering before weighting.

Long prompts, chunks and BREAK

The CLIP text encoder used by Stable Diffusion 1.5 and SDXL reads 77 tokens at a time, of which 75 are usable. Interfaces such as AUTOMATIC1111 handle longer prompts by splitting them into 75-token chunks, encoding each, and joining the results. A chunk boundary can fall in the middle of a phrase, which weakens it.

BREAK (in capitals) forces a new chunk at that point, padding the current one. It is useful for keeping related descriptions together and reducing colour or attribute bleed between subjects:

a portrait of a woman in a yellow raincoat, freckles, wet hair BREAK
a grey concrete underpass behind her, blue fluorescent light BREAK
35mm photo, shallow depth of field, muted colours

It isn’t a guarantee, and each chunk still influences the whole image, but it often keeps “yellow” on the coat rather than spreading it across the wall. In ComfyUI, the equivalent is usually done with separate conditioning nodes that are concatenated or combined.

Tag prompts vs sentence prompts

How much weights matter depends on how the model was trained.

Tag-style models

Many community fine-tunes, especially anime and illustration models, were trained on comma-separated tags. Prompts for them look like this:

1girl, solo, silver hair, (red scarf:1.2), winter coat, snowy street, night, lantern light, looking at viewer

Weights are very effective here, because each tag maps to a well-defined concept the model has seen thousands of times. These models also tend to respond to quality tags, which belong in the positive or negative field depending on the model’s instructions.

Sentence-style models

Base SDXL and many photographic fine-tunes were trained on natural-language captions. They respond better to descriptive phrases:

a young woman with silver hair and a bright red scarf, standing on a snowy street at night, warm light from a paper lantern, looking at the camera

Weights still work on these models, but they matter less, and clear phrasing does more. Weighting a whole phrase, such as (bright red scarf:1.2), works better than weighting single words in the middle of a sentence.

Models where the syntax doesn’t apply

Weighting with brackets is a Stable Diffusion convention. Flux, DALL-E and Gemini treat brackets as ordinary text in most interfaces, so emphasis there comes from word order and explicit description. Midjourney has its own multi-prompt syntax using ::, covered in our Midjourney parameters guide. The Flux vs Midjourney comparison shows how emphasis translates to sentence-based models.

Weights in the negative prompt

Weights work in the negative field the same way. (watermark:1.3) pushes harder against watermarks than watermark. Keep negative weights modest too: a heavily weighted negative can strip out colour or texture along with the thing you wanted removed. Our guide to negative prompts covers which negative terms are worth having at all.

Worked example

Starting prompt, where the lighthouse keeps getting lost:

stormy coastline at dusk, huge waves, dark clouds, dramatic sky, lighthouse, painterly, moody

Step 1, reorder:

a lighthouse on a rocky headland, stormy coastline at dusk, huge waves, dark clouds, painterly, moody

Step 2, add a moderate weight only if needed, and calm the competing element:

a (lighthouse:1.2) on a rocky headland, its beam cutting through the rain, stormy coastline at dusk,
(huge waves:0.9), dark clouds, painterly, moody

Step 3, if the beam now turns everything yellow, that is concept bleed. Lower the weight or move the light description into its own BREAK chunk.

The Stable Diffusion prompt generator produces prompts in this ordered, lightly weighted style from an uploaded image, which makes a good base to tune.

Checklist

1. Put the most important subject first before reaching for weights.
2. Prefer explicit numbers, (word:1.2), over stacked brackets.
3. Stay between 0.8 and 1.3 for most adjustments; treat 1.5 as a ceiling.
4. Weight phrases, not isolated adjectives.
5. Watch for bleed, repetition and burned colour, and lower the weight first.
6. Use BREAK to separate subjects that keep swapping colours or attributes.
7. Match the style to the model: tags for tag-trained models, sentences for caption-trained ones.
8. Retune weights when moving between interfaces.

For more on writing the descriptive part well, see seven habits that fix generic AI images.