Diffusion Docs

Appearance

Prompt Structure ​

An effective prompt gives the model useful conditioning without burying the important parts.

Quick Take

Put the subject and scene first. Add style, lighting, camera, medium, or constraints after the core idea is clear.

A Practical Structure ​

text
subject, action or pose, environment, composition, lighting, medium, style details

Example:

text
a blue ceramic teapot on a wooden kitchen table, morning window light, simple background, shallow depth of field

What To Add Carefully ​

  • many competing styles
  • long lists of adjectives
  • vague quality words
  • terms copied from another model family
  • prompt syntax from another UI

Make The Main Idea Easy To Find ​

The model receives encoded conditioning, not a neatly parsed outline. Still, clear prompt order helps you write and debug. Put the subject and scene before decorative details.

Weak structure:

text
masterpiece, intricate, cinematic, beautiful, dramatic, award winning, blue ceramic teapot

Clearer structure:

text
a blue ceramic teapot on a wooden kitchen table, morning window light, simple background, soft ceramic glaze

Use Specific Nouns And Relationships ​

Specific nouns and spatial relationships usually help more than generic quality words:

  • "one glass vase on a stone windowsill" is clearer than "beautiful object"
  • "three-quarter portrait, looking left" is clearer than "dynamic composition"
  • "small cabin beside a frozen lake" is clearer than "epic landscape"

Revise By Failure Type ​

If the subject is missing, simplify and move it earlier. If the style is wrong, reduce competing style phrases. If the composition is wrong, use clearer framing or aspect ratio. If details are wrong but the composition is good, consider inpainting instead of rewriting the whole prompt.

Prompts Compete For Attention

Long prompts can work, but every extra phrase adds another direction. If a prompt becomes unreliable, shorten it and reintroduce details one at a time.