Stable Diffusion Prompt Order Matter

  • Negative prompts also have order priority, with words at the end of the negative prompt list having higher priority to be excluded from the generated image ([6]).
  • When iterating on a prompt, adding descriptive details at the end can help the AI remember what to emphasize (e.g. “a tall Hispanic man wearing a black leather jacket, 1950s diner aesthetic”) ([13]).
  • Style keywords like “impressionism” or “Disney style” should generally go at the end of the prompt after describing the main subject, so the style can be applied to the key subject elements described earlier in the prompt ([10]).
  • There is a token limit to prompts in Stable Diffusion, usually around 75 tokens, though some services allow going over this. The AI converts text to tokens, which are numerical representations of words, so unfamiliar words may take up more tokens ([16]).
  • Various advanced prompt operators like (), [], ;, : can also control order and emphasis of words in a prompt ([17]).

So in summary, prompt order significantly impacts the AI’s interpretation and creative choices. Putting the most important words and styles first, adding extra descriptive details later, and using operators to control order and emphasis gives the best and most consistent results.