Skip to main content
AI image generation includes randomness. A good prompt improves the odds of the result you want; it does not make every generation identical.

​
Match the prompt style to the model

For Dreamy and Fluidity models, short visual tags are usually more reliable than a paragraph of backstory. Instead of:
They are a lean athletic person with a warm complexion who usually wears their hair up while exercising.
Try:
Vivid and image-editing models can respond well to more natural descriptions. Whichever model you use, describe what should be visible in the image. The image model cannot apply conditional biography such as “sometimes wears” or “only when exercising.”

​
Use a focused structure

A practical order is:
Keep related details together:
is usually clearer than:
Long prompts create more opportunities for conflict. As a starting point, aim for roughly 50 to 60 useful words and remove anything that does not change the visible result.

​
Prompt weights

Parentheses increase emphasis. A numeric weight gives more precise control:
Use small adjustments: Weights that are too high often create distortion or make output less consistent. Start at normal strength, move by 0.1, and stop as soon as the trait is reliable. For Community Presets, keep weights at 1.1 or below in most cases.

​
What to leave out

Avoid inputs the model cannot use reliably:
  • Hex colors, RGB values, exact heights, weights, cup sizes, or body measurements.
  • Uncommon technical terms when a common visual description exists.
  • Quality tags such as masterpiece, best_quality, 4k, or scoring tags.
  • Story logic, smells, sounds, or feelings that have no visible expression.
  • Negative phrasing inside the positive prompt, such as “no jacket.”
Use the dedicated Negative Prompt field to describe unwanted visual elements instead of adding “no” to the positive prompt.

​
Anime Fluidity

Fluidity responds especially well to exact Danbooru-style visual tags. If a common term is inconsistent, use the exact spelling of a relevant tag from a tag library.
  • Exact multiword tags use underscores, for example looking_at_viewer.
  • If a phrase is not an exact tag, use normal spaces rather than inventing an underscored version.
  • Fluidity tends toward dramatic anatomy and flatter color. You may need subtler body presets, while colorful_palette or anime_coloring can strengthen color.
The Anime Default model can still be useful for sketch-like styles, certain fantasy anatomies, or common words that Fluidity does not reproduce well.

​
A reliable tuning loop

  1. Remove optional custom prompts.
  2. Generate a baseline.
  3. Add one visual instruction.
  4. Generate again.
  5. If it had no effect, remove it or try a common synonym.
  6. If it was weak, raise the weight slightly.
  7. If it introduced artifacts, lower the weight or simplify the phrase.
  8. Repeat until the character is close, then move one-off scene details into the Generate page.
The Remix or More → Details action on an existing image lets you inspect the prompt used to generate it.
Last modified on July 27, 2026