Stable Diffusion token weight syntax builder
Emphasising the right words is how you get Stable Diffusion to actually paint
what you asked for — but hand-typing ((nested)) brackets is error-prone and a
single mismatched parenthesis can break the whole prompt. Paste your prompt, set
an exact weight per token with a slider, and copy correct
(token:weight) syntax with no typos.
How attention weighting works
Stable Diffusion converts your prompt into token embeddings and lets you scale the attention each token receives:
(token:1.3) exactly 1.3× attention (emphasis)
(token) ≈ 1.1× attention
[token] ≈ 0.9× attention (de-emphasis)
((token)) ≈ 1.21× (nesting multiplies)The legacy nested-bracket forms are equivalent to a single numeric weight, which
is why this builder always emits the explicit (token:weight) form — it is the
clearest and works identically across AUTOMATIC1111, Forge and ComfyUI. A weight
of exactly 1.0 is left as plain text, since multiplying by one changes nothing.
Why mismatched brackets break prompts
AUTOMATIC1111’s prompt parser uses parentheses to identify weighted regions. An opening ( without a closing ) — or a closing ) without a matching open — causes the parser to misread the rest of the prompt. In the best case it ignores the malformed section; in the worst case it applies the wrong weight to a large portion of your prompt, producing results that look nothing like what you intended. With nested brackets the risk compounds: (((red dress))) means 1.1 × 1.1 × 1.1 = 1.331×, and one missing bracket changes the entire calculation.
The explicit numeric form (red dress:1.3) is unambiguous: the weight is stated, the brackets are a single matched pair, and the parser cannot misread it. This builder always produces the numeric form so you never need to count nesting levels.
Practical examples
| Goal | Plain text | With weight |
|---|---|---|
| Emphasise the subject | a woman in red | (a woman in red:1.3) |
| Reduce background influence | forest background | (forest background:0.7) |
| Boost lighting token | dramatic side lighting | (dramatic side lighting:1.2) |
| Soften an overriding element | clouds | (clouds:0.8) |
Notice that not everything needs weighting. Most tokens in a prompt at 1.0 is a clean and coherent result; weight only the tokens that are being ignored or are dominating unexpectedly.
Tips for clean weighting
- Stay between 0.8 and 1.4. This is the sweet spot for most checkpoints.
- Boost one thing at a time. Weighting everything to
1.3is the same as weighting nothing — you need contrast between weighted and unweighted tokens. - De-emphasise instead of deleting. If a concept keeps dominating, try
0.8–0.9before removing it entirely; removing it entirely can cause the model to substitute something unexpected. - SDXL is gentler. It responds to smaller nudges, so start lower than you would on SD 1.5 and step up incrementally.
- Test one weight change at a time to understand which token is driving a change in the output — changing multiple weights at once makes it hard to attribute the result.