Prompt weighting & LoRA syntax: the complete guide

Attention weights and LoRA syntax are how you tell an image model what matters most. Here's how (token:1.2), nested parentheses, <lora:name:0.8> and BREAK actually work — and how they differ between Automatic1111 and ComfyUI.

What is prompt weighting?

Every word in a prompt competes for the model's attention. Weighting lets you turn the dial up or down on specific tokens so the model emphasises — or ignores — them. A weight above1.0 strengthens a concept; below 1.0 weakens it.

Numeric weights: (token:1.2)

The clearest syntax is an explicit multiplier. Wrap the text in parentheses and add a colon plus a number:

  • (detailed eyes:1.3) — about 30% more emphasis.
  • (background:0.6) — pushed well into the background.
  • 1.0 is neutral; most useful values sit between 0.7 and 1.5. Past ~1.6 images tend to distort.

Bracket emphasis: (((this))) and [[this]]

Automatic1111 also supports bare brackets. Each layer of ( ) multiplies a token's weight by roughly 1.1, and each [ ] divides by roughly 1.1. So ((vivid)) ≈ 1.21× and[muted] ≈ 0.91×. The numeric form is easier to reason about, so prefer it once you're past one or two layers.

LoRA syntax: <lora:name:weight>

A LoRA is a small add-on model that injects a style, character or concept. In Automatic1111 and Forge you load it inline:

  • <lora:add-detail:0.8> — load the add-detail LoRA at 0.8 strength.
  • Typical strengths are 0.6–1.0; stack multiple LoRAs by adding more tags.
  • Many LoRAs also need trigger words in the prompt to activate — check the model card.

In ComfyUI you usually apply a LoRA with a Load LoRA node instead of inline text, with separate model and CLIP strength sliders.

BREAK: isolating concepts

CLIP reads prompts in chunks of 75 tokens. Putting BREAK between two groups starts a fresh chunk so they don't bleed into each other — handy when colours or attributes are leaking across subjects. ComfyUI achieves the same with Conditioning (Concat) nodes.

Automatic1111 vs ComfyUI: what carries over

  • Works in both: comma-separated tags and the (token:1.2) numeric form.
  • A1111 / Forge only: nested ( ) / [ ] emphasis and inline <lora:…>.
  • ComfyUI: LoRAs and conditioning go through nodes; positive and negative prompts are separate inputs.

See the companion ComfyUI & Automatic1111 prompt guide for engine-specific tips.

Frequently asked questions

What does (token:1.2) mean in a prompt?

It multiplies the attention weight of “token” by 1.2, making the model emphasise it about 20% more. Values below 1.0 de-emphasise it; 1.0 is neutral.

How do I weight a prompt in Automatic1111?

Wrap the text in parentheses with a number — (detailed eyes:1.3) — or use bare parentheses, where each ( ) layer multiplies weight by ~1.1 and each [ ] divides by ~1.1.

What is LoRA syntax?

A LoRA is loaded inline with <lora:filename:weight>, for example <lora:add-detail:0.8>. The weight scales the LoRA’s strength; 0.6–1.0 is typical.

Does ComfyUI use the same weighting as Automatic1111?

ComfyUI supports the (token:1.2) numeric form, but not A1111’s nested ( ) / [ ] multipliers by default, and LoRAs are usually applied with a Load LoRA node rather than inline <lora:…> text.

Build it without the syntax wrangling

Prompt Builder turns weighting, LoRA tags and negative prompts into one-click snippets and keyboard shortcuts — free on web, macOS, Windows and Linux.