Prompting

Stable Diffusion Prompt Weights Explained with Mobile-Friendly Examples

PhoneDiffusion
Stable Diffusion Prompt Weights Explained with Mobile-Friendly Examples

Prompts

Stable Diffusion prompt weights tell the model how much attention to give a word. In AUTOMATIC1111, (word) multiplies it by 1.1, ((word)) by 1.21, (word:1.4) sets an exact weight and [word] lowers it. ComfyUI and the Diffusers helper libraries use similar syntax but apply it differently, so start with small steps such as (word:1.2) and compare with a fixed seed.

Prompt weight syntax in AUTOMATIC1111

The AUTOMATIC1111 wiki calls this attention or emphasis. Parentheses raise the attention a word or phrase gets and square brackets lower it. (fog) multiplies it by 1.1, and nesting compounds the effect: ((fog)) is 1.1 × 1.1, or 1.21. [fog] divides it by 1.1.

For an exact value, write the number after a colon inside parentheses: (fog:1.4) raises it, and (fog:0.25) lowers it by a factor of 4. The wiki notes that a number only works with parentheses, not with square brackets. To use real parentheses in a prompt, for example in a tag name, escape them with a backslash, as in anime_\(character\).

How weights differ in AUTOMATIC1111, ComfyUI and Diffusers

AUTOMATIC1111 has an Emphasis mode setting. The default, Original, multiplies each token's encoding by its weight and then rescales the result so its average stays the same; No norm skips that rescaling, and the setting describes it as seeming to work better for SDXL. None turns the syntax off and reads the brackets as plain characters. Because the mode changes the result, the same prompt and seed can give a different image when the mode differs.

ComfyUI's README documents the same parentheses: (good code:1.2) or (bad code:0.8), with 1.1 as the default for plain (). Two differences matter. ComfyUI's parser only reads parentheses, so [word] is not a way to lower a weight there. And its code applies a weight by scaling how far a word's encoding sits from the encoding of an empty prompt, not by AUTOMATIC1111's multiply-then-rescale. A prompt copied between the two tools can therefore look different at the same numbers.

In Diffusers, weights do not go in the plain prompt string. The prompting guide applies them through prompt_embeds, built with a helper library. sd_embed reads the familiar syntax: (cat) is 1.1, ((cat)) 1.21 and (cat:1.5) 1.5. Compel has its own: cat+ is 1.1, cat++ is 1.1², cat- is 0.9, and a number goes after the closing parenthesis, as in (picking apricots)1.3. Compel lowers a word by blending with a version of the prompt without it, rather than shrinking the word directly.

Sensible weight ranges

No document gives one correct range, because the effect depends on the model, the tool and the rest of the prompt. The examples maintainers use stay close to 1: AUTOMATIC1111's setting text shows (more:1.1) and (less:0.9), ComfyUI's README uses 1.2 and 0.8, and Compel accepts weights between 0 and 2.

A practical way to work: start at 1.1 or 1.2, change one weight at a time, and keep the seed, steps, sampler and size fixed so the weight is the only thing that changed. Stable Diffusion seeds explained shows how to lock the seed. If a word still does not show at a moderate weight, try rewording it or moving it earlier in the prompt before reaching for a bigger number.

PhoneDiffusion generates on-device on iPhone or iPad, with optional cloud models. Download on the App Store →

Mobile-friendly examples

These are short enough to type on a phone and use only the parentheses and colon on the number keyboard. They are AUTOMATIC1111-style syntax, which ComfyUI and sd_embed also read.

‘portrait of an old fisherman, (weathered skin:1.2), soft window light’ strengthens one detail without touching the rest. ‘a red bicycle on a (foggy:1.3) street at dawn’ pushes the mood. ‘cozy cafe interior, warm light, (people:0.7)’ turns down the attention on people; a negative prompt is the other way to keep something out. In Compel, the first example becomes ‘portrait of an old fisherman, (weathered skin)1.2, soft window light’.

If you draft prompts in a notes app first, check the punctuation before you paste. Apple's Smart Punctuation turns a double hyphen into a dash and straight quotation marks into curly ones as you type, and the setting applies to other apps too. Compel needs plain hyphens for cat-- and straight double quotes around the parts of .blend() and .and(), so those changes stop the syntax from working. Turn it off in Settings > General > Keyboard > Smart Punctuation, or retype those characters on the desktop.

Short prompts are also easier to reason about. The AUTOMATIC1111 wiki explains that prompts are processed in 75-token chunks, each encoded separately, and that the uppercase BREAK keyword ends the current chunk so the next words start a new one.

Common mistakes with prompt weights

Putting a number in square brackets. In AUTOMATIC1111, [word:1.2] is not a weight: square brackets with colons are its prompt-editing syntax, as in [fantasy:cyberpunk:16], which switches words partway through the steps. Use (word:1.2) for a weight.

Copying a prompt between tools unchanged. [word] does nothing special in ComfyUI, Compel expects (word)1.2 instead of (word:1.2), and AUTOMATIC1111's Emphasis mode changes how strong a weight looks. Convert the syntax and retest when you move a prompt.

Forgetting to close or escape parentheses. An unclosed bracket or a tag such as anime_(character) written without backslashes gets read as weighting. Escape literal brackets as \( and \).

Mixing up prompt weights and LoRA weights. <lora:name:0.7> sets how strongly a LoRA changes the model, which is a different control; see Stable Diffusion LoRA weights explained.

Raising several weights at once. When three words go up together, you cannot tell which change helped. Move one, compare, then move the next.

PhoneDiffusion availability and article scope

PhoneDiffusion does not parse weight syntax such as (word:1.2), so parentheses and numbers are not a way to strengthen a word in the app. Write the important detail in plain words near the start of the prompt, and use the optional negative prompt in local generation for things you want to keep out. The curated local Core ML packs run on supported iPhone and iPad devices with iOS 17 or later, privately and offline once downloaded.

Optional cloud models, such as FLUX and Seedream, on paid monthly or yearly plans are text-to-image only and send the text prompt online, without a negative prompt; plain descriptive wording is the safer choice there too. The weight syntax, Emphasis modes, Compel and sd_embed described above belong to AUTOMATIC1111, ComfyUI and Diffusers on desktop or server.

Field notes

  • In AUTOMATIC1111, (word) is 1.1×, ((word)) is 1.21×, (word:1.4) is an exact weight and [word] is weaker.
  • Numbers only work inside parentheses; square brackets with colons are prompt editing.
  • ComfyUI reads parentheses but not [word], and applies weights differently from AUTOMATIC1111.
  • Diffusers needs a helper: sd_embed reads the same syntax, Compel uses + and - or (phrase)1.3.
  • Start at 1.1 or 1.2 and change one weight at a time with a fixed seed.

Try it on your phone

  1. Write the prompt in plain words first and generate a baseline with a fixed seed.
  2. Pick the one word that is too weak and wrap it, for example (foggy:1.2).
  3. Generate again with the same seed and settings, then adjust in small steps.
  4. Lower a distracting word with (word:0.8) or move it to the negative prompt.
  5. Convert the syntax before moving the prompt to another tool, and save the final weights with the image.

Sources and verification

Primary and maintainer documentation checked for this guide. Product-specific details were reviewed against the current PhoneDiffusion scope.

  • AUTOMATIC1111 web UI Features wiki: attention/emphasis, prompt length, negative promptProject wiki: (word) raises attention by 1.1, ((word)) by 1.21, [word] lowers it by 1.1, (word:1.5) sets a weight; prompts are split into 75-token chunks; the negative prompt replaces the empty unconditional prompt during sampling.
  • AUTOMATIC1111 web UI: Emphasis mode options (sd_emphasis.py, shared_options.py)Source for the Emphasis mode setting: Original multiplies each token's encoding by its weight and then restores the original mean; No norm skips that step (‘seems to work better for SDXL’); Ignore drops all emphasis; None reads the brackets as literal characters. A separate option uses the old emphasis implementation to reproduce old seeds.
  • ComfyUI: README prompt syntax and the CLIP weight code (comfy/sd1_clip.py)README: (good code:1.2) or (bad code:0.8) changes emphasis, the default for () is 1.1, escape literal parentheses as \( and \), {day|night} is a random wildcard. Source: only parentheses are parsed for weights, and a weighted token is encoded as (z − z_empty) × weight + z_empty, measured from an empty prompt.
  • Hugging Face Diffusers: Prompting (guidance, negatives, prompt weighting)Maintainer guide: build prompts from subject, style and context; negative_prompt and guidance_scale on Stable Diffusion pipelines; prompt weighting through prompt_embeds, with sd_embed reading (cat), ((cat)) and (cat:1.5), or with Compel, which the guide describes as a simpler syntax.
  • Compel: prompt weighting syntax (doc/syntax.md)Maintainer doc: append + or - or a weight between 0 and 2; + is 1.1, ++ is 1.1², - is 0.9, -- is 0.9²; (picking apricots)1.3 sets a number for a group; nested modifiers stack; downweighting is a weighted blend with the prompt without the term; .blend() and .and() take quoted prompt parts.
  • Apple Pages User Guide for iPhone: Smart PunctuationSmart Punctuation converts double hyphens to a dash and straight quotation marks to curly ones as you type; turn it off in Settings > General > Keyboard > Smart Punctuation; the setting applies to other apps you use to enter text.