Stable Diffusion prompt weights adjust the influence of selected words: use (word:1.2) to increase emphasis or (word:0.8) to reduce it. Plain text has weight 1.0, while (word) defaults to 1.1 in A1111, ComfyUI, and Fooocus.
Square brackets, prompt scheduling, and the underlying weighting calculations differ by interface, so use the syntax table before copying a prompt between tools. Sources: A1111 parser, ComfyUI documentation, Fooocus parser.
What is Prompt Weighting?
Prompt weighting changes how parts of your description contribute to the text conditioning used for generation. The interface parses your syntax before the model receives that conditioning. Sources: A1111 emphasis, ComfyUI encoder.
In a prompt such as woman, beach, pizza, unweighted text receives a multiplier of 1.0. This does not mean each word owns one-third of the picture or one-third of the model's attention. Sources: parsers, ComfyUI encoder.
The weights are controls on encoded text, not percentages of objects or guaranteed screen space. A larger coefficient does not specify the size, location, or count of the named object.
If you are choosing an interface, our AI image generator comparison provides broader context.
Increasing word weights
Start with an unweighted description and change one phrase:
woman eating pizza on a beach
woman eating pizza on a (beach:1.2)
The second prompt assigns the beach phrase a higher weight. Compare framing, background, and subject placement; do not judge the result solely by whether a beach appears. The syntax follows the CLIP Text Encode documentation.
Compare the illustrated compositions, then repeat the experiment in your interface with the model and generation settings held constant.
Every word counts
A longer prompt does not divide attention into equal shares. A word can span multiple tokens, and its encoded representation includes context from surrounding text. Source: Advanced CLIP documentation.
Try a more specific description:
photograph of a woman with dark hair eating pepperoni pizza
on a (white beach:1.2), sunset, ocean, natural light
Here the modifier covers white beach together. Keep the rest of the description unchanged during the comparison so that you can assess the contribution of that phrase.
If the composition becomes less useful, simplify the wording or remove competing details before increasing the coefficient again.
Syntax by UI
This table describes standard A1111 emphasis and the built-in Stable Diffusion CLIP text paths in ComfyUI and Fooocus. Custom nodes or extensions can introduce different parsers.
| Interface | Increase emphasis | Reduce emphasis | Square brackets |
AND / [a:b:0.5]
|
|---|---|---|---|---|
| AUTOMATIC1111 |
(word) = 1.1; (word:1.3) is explicit |
(word:0.8) |
[word] = 1/1.1 |
Supported as separate composition / scheduling features |
| ComfyUI built-in CLIP |
(word) = 1.1; (word:1.2) is explicit |
(word:0.8) |
No decrement; treated as text | Not interpreted as those operators by the built-in text encoder |
| Fooocus |
(word) = 1.1; (word:1.2) is explicit |
(word:0.8) |
No decrement in its parser | Not supported as A1111 composition / scheduling operators in the standard path |
A1111's advanced operators are documented in its Features wiki. Fooocus's generation worker processes prompts without A1111's composition or step-scheduling parser.
How do I use it?
Choose the smallest phrase that describes your intended change. Generate a baseline, apply one explicit weight, and compare while keeping the checkpoint, seed, dimensions, sampler, steps, and other prompt text unchanged.
Try another seed before adopting the change. Evaluate each coefficient against your intended result.
Prompt Weight in Automatic1111
A1111 supports both emphasis shortcuts and explicit coefficients. Its default Emphasis mode is Original; other modes can change or disable weighting behavior. Source: settings definitions.
-
(blue eyes)applies a 1.1 multiplier. -
((blue eyes))applies 1.1 × 1.1, or 1.21. -
(blue eyes:1.3)supplies the coefficient directly. -
[blue eyes]applies 1/1.1, approximately 0.9091. -
(blue eyes:0.8)explicitly reduces the weight to 0.8.
These values follow the A1111 attention parser. Square-bracket reduction is not exactly 0.9. Use an explicit numeric weight when that distinction matters to your comparison.
Numeric weights belong inside parentheses. [blue eyes:0.8] is not the documented numeric emphasis form and can be interpreted as prompt editing. Source: A1111 Features.
For literal parentheses, escape them:
portrait of a character \(adult\), blue jacket
The backslashes preserve the parentheses as text instead of starting an emphasis group. A1111 also accepts escaped square brackets, \[ and \]. Source: A1111 parser.
AND combines prompt conditionings
Uppercase AND is A1111's composable-diffusion operator. It separates prompts, and each part can have its own coefficient:
a lighthouse:1.2 AND a stormy coastline:0.8
This combines separately conditioned prompts; it is not equivalent to emphasizing one phrase in a single prompt. It also does not assign the lighthouse to a particular region of the canvas. Source: composable diffusion documentation.
Prompt editing changes text during sampling
For A1111's initial sampling pass:
a [red:blue:0.5] ceramic bowl on a table
The prompt uses red for the first half of the steps and blue afterward. The value 0.5 specifies a switching point, not a weight or an equal color mixture. Source: scheduling implementation.
Integer timing values specify steps; decimals specify fractions. Hires-fix scheduling has a separate timeline, so check its rules before transferring a first-pass schedule. Source: A1111 prompt editing.
Prompt Weight in Fooocus
Fooocus supports (phrase:weight), plain parentheses, and escaped literal parentheses. Its README explicitly documents weighted phrases. Sources: README, parser.
For a reduced weight, use (phrase:0.8). Its parser does not implement A1111's square-bracket decrement, even though the README describes its reweighting as A1111-compatible. Sources: parser, README.
Select text and use Ctrl+Up or Ctrl+Down to edit attention. Inspect the resulting text before generating. Source: change log.
Keep styles fixed during a weighting comparison. Fooocus can add style text and prompt expansion, so a changed style selection introduces another difference alongside your coefficient. Sources: README, generation worker.
Prompt Weight in ComfyUI
In ComfyUI, enter (phrase:1.2) in the CLIP Text Encode node for increased emphasis, or (phrase:0.8) for reduced emphasis. Plain (phrase) means 1.1. Source: node documentation.
The built-in parser does not use [phrase] for reduction. It also does not turn A1111's AND or [a:b:0.5] into composition or scheduling operations. Those require an appropriate workflow or extension. Source: encoder implementation.
You can select prompt text and press Ctrl+Up/Down to adjust its weight. The adjustment size is configurable. Escape literal parentheses with \( and \). Source: node documentation.
Advanced CLIP Encode weight_interpretation
BlenderNeko's CLIP Text Encode (Advanced) is a custom node, separate from the built-in encoder. Its weight_interpretation options are comfy, A1111, compel, comfy++, and down_weight. Source: extension README.
Its separate token_normalization setting offers none, mean, length, and length+mean. Record both settings when sharing a workflow; selecting an interpretation changes how the node handles the weights. Source: extension README.
Difference between Automatic1111 and ComfyUI
A1111's Original emphasis multiplies encoded token vectors by their coefficients, then rescales the result to restore the original mean. It does not normalize word weights so they sum to one. Source: emphasis implementation.
ComfyUI's built-in encoder instead applies weighting relative to an empty-prompt embedding. In simplified form, it computes empty + weight × (token − empty) for a weighted token. Source: ComfyUI encoder.
Fooocus uses that empty-prompt-relative operation, then restores the original mean. Its README's compatibility wording does not establish identical computations or images. Source: Fooocus CLIP patch.
These implementation differences mean shared syntax does not establish equivalent encodings. Treat a copied weight as a starting point to evaluate in the destination workflow.
Worked examples
Emphasize a complete attribute
portrait of a woman wearing a (blue linen jacket:1.2), daylight
Compare with the unweighted prompt. Inspect whether the jacket, color, and overall composition remain useful.
Reduce emphasis without removing the concept
mountain landscape, (mist:0.8), pine forest, morning light
Compare the reduced-weight version with mist unweighted. Down-weighting changes conditioning; it does not remove the word from the description. If you do not want mist at all, first try deleting that phrase. Source: encoder.
Separate timing from strength
studio photograph of a [glass:ceramic:0.5] vase
Use this only with A1111's scheduling support. It asks for an early glass-conditioned phase followed by a ceramic-conditioned phase; it does not assign either material a 0.5 emphasis weight. Source: scheduler.
Common mistakes
-
Copying square brackets into ComfyUI or Fooocus: use explicit
(phrase:0.8)reduction for the documented built-in paths. - Weighting everything: choose the phrase you want to investigate so that the comparison stays interpretable.
- Changing several controls together: hold other settings fixed while evaluating a coefficient.
- Expecting object placement or percentages: token weights do not define layout regions or image-area shares.
- Assuming nested numeric groups transfer exactly: use flat explicit groups when moving prompts between different parsers.
- Ignoring the active encoder: custom-node interpretation and normalization settings can change the result.
The syntax cautions follow the A1111, ComfyUI, and Fooocus parsers. The Advanced CLIP README explains why encoder settings matter.
FAQ
What is the default Stable Diffusion prompt weight?
Unweighted text uses 1.0 in these parsers. Plain parentheses apply 1.1; an explicit coefficient such as (word:1.2) selects that value. Sources: A1111, ComfyUI, Fooocus.
Do brackets decrease prompt weight?
In A1111, [word] multiplies by 1/1.1. ComfyUI and Fooocus do not implement that decrement. Use (word:0.8) for explicit reduction across all three. Sources: A1111, ComfyUI, Fooocus.
What prompt weight gives the best results?
No value is universally best. Compare an unweighted phrase with a modest explicit change, and choose according to your intended image. Keep the interface, encoder configuration, and other generation settings consistent during the test.
Is [a:b:0.5] the same as (b:0.5)?
No. In A1111, the first schedules a change from a to b halfway through the initial sampling pass. The second reduces the weight of b throughout its use in the prompt. Source: A1111 parser.
Sources
- AUTOMATIC1111 Features: emphasis, composition, prompt editing
- AUTOMATIC1111 attention and scheduling parser
- AUTOMATIC1111 emphasis calculations
- AUTOMATIC1111 default emphasis setting
- ComfyUI CLIP Text Encode documentation
- ComfyUI parser and token-weight encoder
- Advanced CLIP Text Encode extension documentation
- Fooocus README
- Fooocus prompt parser
- Fooocus CLIP weighting implementation
- Fooocus attention keyboard shortcuts
- Fooocus generation and prompt-processing path
Last reviewed and updated: September 2026.
Top comments (2)
Great breakdown of prompt weighting! I've been playing around with Stable Diffusion but always felt like my results were hit or miss. This explains why some keywords seemed to get "lost" in longer prompts. Definitely going to try adjusting weights in my next few attempts. Thanks for the clear examples and interface comparisons too - super helpful!
Great article! It explains how to adjust prompt weights in Stable Diffusion to improve image generation. The comparisons between different interfaces and practical tips are really helpful. Thanks for the insights!