How does Stable Diffusion weighting actually work?

In the AUTOMATIC1111 / SD WebUI convention, each pair of round parentheses around a token multiplies its attention weight by 1.1, and each pair of square brackets multiplies it by 0.9. Nesting compounds, so ((token)) is 1.1 × 1.1 ≈ 1.21. The cleaner form is the explicit (token:1.21), which avoids stacking brackets.

What's the difference between (token) and (token:1.5)?

"(token)" is shorthand for "(token:1.1)". The explicit "(token:1.5)" lets you set any multiplier directly and is far easier to read than nesting five sets of parentheses. Most experienced users prefer the explicit numeric form.

How high or low should I push a weight?

Useful emphasis usually lives between about 0.7 and 1.5. Going above ~1.4 often distorts the image or causes the token to dominate unnaturally; going below ~0.6 may make the token effectively disappear. Adjust in small steps and watch the result.

Does this apply to every model and UI?

This 1.1 / 0.9 parenthesis convention is the AUTOMATIC1111 / Stable Diffusion WebUI standard, also followed by many forks and ComfyUI nodes. Some tools (including Midjourney and DALL·E) use different or no weighting syntax, so check your tool's documentation.

SD Prompt Attention Weight Calculator

SD prompt attention weight calculator

In Stable Diffusion’s WebUI convention, you emphasise a token by wrapping it in parentheses and de-emphasise it with square brackets — and the effect compounds when you nest them. This calculator turns “how many brackets did I stack?” into an exact multiplier, and gives you the cleaner explicit syntax to use instead.

How it works

Each pair of round parentheses multiplies a token’s attention weight by 1.1; each pair of square brackets multiplies it by 0.9. Nesting multiplies again:

(token)      → 1.1
((token))    → 1.1 × 1.1 ≈ 1.21
(((token)))  → 1.1³       ≈ 1.331
[token]      → 0.9
[[token]]    → 0.9 × 0.9 = 0.81

So the effective weight for depth n is 1.1ⁿ for emphasis or 0.9ⁿ for de-emphasis. The calculator computes that and shows the equivalent (token:weight) form, which is far easier to read than a pile of brackets.

Tips and notes

Prefer the explicit form. (detailed face:1.3) is clearer and less error-prone than ((detailed face)).
Stay in a sane range. Emphasis above ~1.4 tends to distort; below ~0.6 the token may vanish. Adjust in small steps.
Weights are relative. Pushing one token up effectively pushes the rest down, so re-balance the whole prompt rather than only inflating one word.
UI matters. This is the AUTOMATIC1111 / SD WebUI standard. ComfyUI and other tools mostly follow it, but Midjourney and DALL·E do not — check your tool.