Prompt weighting controls how strongly each part of a prompt is applied. Use (word:1.3) to boost attention, ((word)) brackets to nudge it, BREAK to separate concepts, and [a:b:0.4] to switch a term partway through sampling. Syntax differs across AUTOMATIC1111, Forge, and ComfyUI, and Flux handles emphasis differently. All examples use adult, fictional, AI-generated characters with baseline safety negatives.
Weighting is the difference between hoping the model notices a detail and telling it exactly how much to care. Once you understand attention syntax, you stop rewriting prompts ten times and start dialing them in precisely. This guide covers every weighting tool that matters for adult AI work, the differences across the major interfaces, and a syntax reference table you can keep open. Throughout, the safety baseline is fixed: every character is an adult (18 and over), fictional, and AI-generated, and every prompt carries the child, minor, underage, loli, shota negative. Weighting changes emphasis, never the rules.
What weighting actually does
In Stable Diffusion family models, the text encoder turns your prompt into a set of token embeddings, and attention weighting scales how much influence each token has during sampling. A weight of 1.0 is neutral. Above 1.0 pushes the model to express that token more strongly. Below 1.0 pulls it back. The practical effect is control: you can make a hard-to-summon detail show up, or tame a token that is taking over the image.
This is most useful for adult work where specific attributes, poses, or styles are competing for the model’s attention. If a build descriptor, an expression, or a lighting term keeps getting ignored, a small weight bump often fixes it without rewriting the whole prompt. If you are still assembling the prompt itself, our NSFW AI prompt formula covers what goes in before you start tuning weights.
Attention syntax: the colon method
The clearest and most common syntax is the parenthesis-colon form: (word:1.3). The number is the weight. This works in AUTOMATIC1111, Forge, and most SD interfaces, and it is the form you should default to because it is explicit and readable.
photo of an adult woman, 27 years old, fictional AI character,
(athletic build:1.2), (detailed eyes:1.1), soft window light, photorealistic
Negative: child, minor, underage, loli, shota, (bad anatomy:1.2), lowres,
deformed, extra limbs, watermark, textYou can weight phrases, not just single words, by wrapping the whole phrase: (soft window light:1.15). You can also weight negative tokens, which is useful for stubborn artifacts, for example (extra fingers:1.3) in the negative.
Keep positive weights in a sane range. Roughly 1.05 to 1.3 is the working zone. Above about 1.4 the token starts to dominate and distort anatomy or color. For de-emphasis, 0.7 to 0.95 gently pulls a token back without removing it.
Bracket and parenthesis shorthand
An older shorthand uses bare brackets. Each pair of parentheses multiplies attention by about 1.1, and each pair of square brackets divides by about 1.1.
((word))is roughly(word:1.21).(word)is roughly(word:1.1).[word]is roughly(word:0.9).
This is quick to type but imprecise and hard to read once you stack several. The colon method is clearer and is preferred for anything beyond a casual tweak. The bracket shorthand mainly survives in older prompt libraries, so it is worth recognizing even if you do not write it.
score_9, score_8_up, 1woman, adult woman, mature, fictional character,
((curvy hourglass figure)), (detailed face), soft lighting, photorealistic
Negative: child, minor, underage, loli, shota, bad anatomy, lowres, watermarkThe BREAK keyword
BREAK (in capitals) tells the interface to start a new chunk of conditioning. Stable Diffusion processes the prompt in segments of 75 tokens, and BREAK forces a clean boundary so concepts in different chunks do not bleed their attributes into each other. It is invaluable when you describe several distinct elements and do not want their colors, materials, or features to merge.
photo of an adult woman, 29 years old, fictional AI character, red dress,
detailed face, looking at viewer
BREAK
modern apartment interior, soft window light, shallow depth of field,
photorealistic
Negative: child, minor, underage, loli, shota, bad anatomy, lowres, watermark, textHere BREAK keeps the dress color from leaking into the interior and vice versa. For multiple separate characters with distinct attributes, BREAK alone is often not enough and you want true regional control. Our Regional Prompter guide covers assigning prompts to canvas regions, which is the proper tool for two adult characters with different builds or outfits.
Prompt editing: switching terms mid-render
Prompt editing changes the prompt partway through the sampling process using the [from:to:when] syntax. The when value is the fraction of steps at which the switch happens.
[a:b:0.4]starts with term a, then switches to term b at 40 percent of the steps.[word::0.5]includes word for the first half, then drops it.[:word:0.5]adds word only in the second half.
This is powerful for blending two looks or for letting the model establish structure before a detail is introduced. A classic use is to set composition early then refine a feature later.
photo of an adult woman, 30 years old, fictional AI character,
[soft smile:confident expression:0.5], detailed eyes, looking at viewer,
soft studio lighting, photorealistic
Negative: child, minor, underage, loli, shota, blank expression, lowres,
deformed, watermark, textUse prompt editing sparingly. It is a precision tool, and overusing it makes prompts hard to reason about. For most cases a clean weight is simpler and more predictable.
Syntax reference table
| Syntax | Meaning | Range / note |
|---|---|---|
(word:1.3) |
Boost attention to weight 1.3 | 1.05 to 1.3 useful |
(word:0.8) |
Reduce attention | 0.7 to 0.95 gentle |
(word) |
Boost by about 1.1 | shorthand |
((word)) |
Boost by about 1.21 | shorthand |
[word] |
Reduce by about 0.9 | shorthand |
BREAK |
New conditioning chunk | prevents attribute bleed |
[a:b:0.4] |
Switch a to b at 40 percent steps | prompt editing |
[word::0.5] |
Drop word after 50 percent | prompt editing |
[:word:0.5] |
Add word after 50 percent | prompt editing |
Try these in our free generator: take a prompt that is ignoring one attribute, add a 1.2 weight to it, and watch the difference a single number makes.
Differences across interfaces
AUTOMATIC1111
The original reference implementation. It supports the full colon syntax, bracket shorthand, BREAK, and prompt editing exactly as described above. Most prompt libraries are written for A1111, so its behavior is the de facto standard.
Forge
Forge is a performance-focused fork of A1111 and keeps the same prompt syntax, so weights, BREAK, and prompt editing all behave the same way. Anything you learn in A1111 transfers directly. Minor differences exist in how some extensions compute weights, but for everyday prompting they are interchangeable.
ComfyUI
ComfyUI uses the same (word:1.3) colon syntax inside its text encode nodes, so the core weighting carries over. BREAK is handled differently because ComfyUI builds conditioning through nodes: instead of typing BREAK you often use separate conditioning nodes and a combine node to achieve the same separation. Prompt editing also has node-based equivalents. The text weights themselves work as expected. Our ComfyUI complete guide shows how conditioning nodes replace the inline keywords.
Flux and natural-language models
Flux is the important exception. It is built around a different text encoder and is designed to follow natural-language descriptions rather than weighted tag soup. The (word:1.3) syntax does not behave the same way and can be ignored or cause artifacts. To emphasize something in Flux, you describe it more prominently in the sentence, repeat the idea, or restructure the sentence so the important element comes first and is described in more detail. Negatives also work differently. Our Flux NSFW guide covers its prompting model in depth, but the headline is simple: write like you are briefing a person, not weighting tags.
| Interface | Colon weights | BREAK | Prompt editing |
|---|---|---|---|
| AUTOMATIC1111 | Yes | Yes | Yes |
| Forge | Yes | Yes | Yes |
| ComfyUI | Yes | Node-based | Node-based |
| Flux | Not standard | No | No |
Practical weighting recipes
Here are common adult-work situations and the weighting move that fixes each. All use adult, fictional, AI-generated characters with the baseline safety negative.
A build descriptor is ignored. Bump it: (athletic build:1.2). See our body type prompts for the descriptor vocabulary.
An expression is too weak. Bump it gently: (soft smile:1.1), (detailed eyes:1.1). Our facial expression prompts cover the terms.
A style is bleeding in. De-emphasize the intruder: (realistic:0.8) on an anime prompt. Our art style prompts explain style bleed.
An outfit color leaks onto the background. Add BREAK between the character and the setting, or use regional prompting.
A stubborn artifact keeps appearing. Weight the negative: (extra fingers:1.3) or (watermark:1.2). Our negative prompt master list has a full set.
Weighting and safety
This cannot be overstated: weighting changes emphasis, never meaning, and it can never be used to weaken the safety baseline. The adult anchor stays in the positive prompt and the child, minor, underage, loli, shota tokens stay in the negative on every render, regardless of any weights you apply. If anything, you can strengthen the safety steering by weighting the negative tokens, for example (child:1.3) in the negative on models where that helps push harder away from any underage output. The rules are fixed: adults only, fictional AI characters only, never a real identifiable person, never a minor or minor-appearing subject. For more on a complete responsible workflow, our photo editing workflow shows where weighting fits alongside ADetailer and upscaling.
How weighting interacts with the negative prompt
Many people only weight positive tokens, but the negative prompt is where weighting often pays off most in adult work. Stubborn artifacts like extra fingers, watermarks, blurry edges, and deformed hands respond well to a weighted negative because you are telling the model precisely how hard to avoid each one. A flat negative list treats every token equally, but a render with a recurring hand problem benefits from (deformed hands:1.3), (extra fingers:1.3) while the rest of the negatives stay at 1.0.
The same logic applies to style negation. If a photoreal image keeps drifting toward an illustrated look, weighting the competing style in the negative, for example (anime:1.2), (cartoon:1.2), pushes the render back toward realism more firmly than an unweighted token would. This pairs directly with the style separation ideas in our art style prompts guide.
Do not over-weight the negative either. The same 1.4 ceiling applies. A negative token pushed too hard can suppress legitimate detail along with the artifact, so a hand negative at 1.6 might start erasing fingers entirely. Stay in the same gentle range you use for positives, and adjust one token at a time so you can see exactly what each change does to the image.
Reading and reusing other people’s prompts
A large part of getting good at weighting is learning to read the prompts other creators share. When you find a strong adult prompt on a model card or in a community gallery, the weights tell you what the author was struggling with. Heavy weights on a build descriptor usually mean the base model resisted that build. A weighted expression token means the face kept coming out flat. Weighted negatives reveal which artifacts the checkpoint is prone to. Reading weights as a record of problems solved makes any shared prompt far more useful to adapt.
When you reuse a prompt, do not assume its weights transfer perfectly to your checkpoint. A weight tuned for one model may be too strong or too weak on another. Start by lowering aggressive weights toward neutral, generate, and only raise them back if the attribute is missing. Our prompt examples library and best prompt generators roundup are good sources of well-built prompts to study and adapt, and the prompt formula explains the structure those weights sit inside.
Whatever prompt you borrow, audit its safety baseline before you run it. Replace or add a clear adult anchor and confirm the child, minor, underage, loli, shota negative is present. Never trust that a shared prompt already meets the safety standard, since many do not. Adults only, fictional AI characters, never a real person, never a minor or minor-appearing subject, every single time.
Common weighting mistakes
Over-weighting. Weights above roughly 1.4 dominate the image, warp anatomy, and oversaturate color. If a token needs a huge weight to show up, the prompt or checkpoint is probably wrong for it. Pull the weight back and fix the underlying prompt instead.
Stacking shorthand. ((((word)))) is unreadable and imprecise. Use the colon form so you and anyone reading the prompt know the exact weight.
Weighting on Flux as if it were SD. The colon syntax is not standard on Flux. Restructure the sentence instead of bracketing tags.
Forgetting BREAK for multi-element scenes. If two described elements keep merging colors or features, a BREAK between them often solves it instantly.
Putting it together
Weighting turns prompting from guesswork into control. Default to the (word:1.3) colon syntax, keep positive weights between roughly 1.05 and 1.3, use BREAK to stop attribute bleed, and reach for prompt editing only when you genuinely need to switch a term mid-render. Remember that ComfyUI does separation through nodes and that Flux wants natural language rather than weighted tags.
Keep the safety baseline fixed on every render: adults only, fictional AI characters only, never a real identifiable person, never a minor or minor-appearing subject, and child, minor, underage, loli, shota in the negative every time, with weighting available to strengthen that steering rather than weaken it. Build a few weighting habits, test them in our generator, and you will spend far less time rerolling and far more time getting exactly the adult image you intended.
Frequently asked questions
What does the number in (word:1.3) mean?
It is the attention weight applied to that word or phrase. A value of 1.0 is neutral, above 1.0 boosts how strongly the model expresses the token, and below 1.0 pulls it back. So (athletic build:1.2) emphasizes that build, while (realistic:0.8) reduces a style that is bleeding in. Keep positive weights roughly between 1.05 and 1.3, since values above about 1.4 tend to dominate and distort anatomy.
What is the difference between (word) and (word:1.3)?
They are two ways to express attention. The bracket shorthand (word) boosts by about 1.1 and ((word)) by about 1.21, while the colon form (word:1.3) states the exact weight. The colon method is clearer and preferred for anything beyond a quick tweak, because stacked brackets become unreadable and imprecise. The shorthand mainly survives in older prompt libraries, so it is worth recognizing even if you write the colon form.
What does the BREAK keyword do?
BREAK, in capitals, forces a new conditioning chunk so concepts on either side do not bleed their attributes into each other. Stable Diffusion processes the prompt in 75 token segments, and BREAK creates a clean boundary. It is ideal for keeping, for example, a dress color from leaking onto the background. For multiple distinct characters, BREAK alone is usually not enough and regional prompting is the proper tool.
How does prompt editing with [a:b:0.4] work?
It switches the prompt partway through sampling. The syntax [a:b:0.4] starts with term a, then swaps to term b at 40 percent of the steps. Variants include [word::0.5], which drops a word after the halfway point, and [:word:0.5], which adds one only in the second half. It is useful for letting the model set structure before introducing a detail, but use it sparingly since it makes prompts harder to reason about.
Does prompt weighting work the same in ComfyUI?
The core colon syntax (word:1.3) works the same inside ComfyUI text encode nodes. The difference is BREAK and prompt editing, which ComfyUI handles through nodes rather than inline keywords. Instead of typing BREAK you use separate conditioning nodes and a combine node to achieve the same separation. The text weights themselves behave as expected, so most of what you learn in AUTOMATIC1111 transfers directly to ComfyUI.
Why does weighting syntax not work in Flux?
Flux uses a different text encoder built to follow natural-language descriptions rather than weighted tags. The (word:1.3) syntax is not standard there and can be ignored or cause artifacts. To emphasize something in Flux, describe it more prominently, put it earlier in the sentence, or add detail to it. Think of prompting Flux like briefing a person in plain language rather than tuning numerical weights on tags.
Can weighting override the safety negatives?
No. Weighting changes how strongly a token is applied, never its meaning, so it can never weaken the safety baseline. The adult anchor stays in the positive and child, minor, underage, loli, shota stays in the negative on every render. If anything, you can strengthen safety steering by weighting the negative tokens higher on models where that helps. The rules are fixed: adults only, fictional AI characters, never a minor or real person.
What weight is too high?
Roughly above 1.4 a token starts to dominate the image, warp anatomy, and oversaturate color. If a token needs a very high weight just to appear, the underlying prompt or checkpoint is probably wrong for it, so fix that instead of cranking the weight. Stay in the 1.05 to 1.3 range for boosts and 0.7 to 0.95 for gentle reductions, which keeps the image balanced and avoids the distortion that extreme weights cause.
