Skip to content Skip to footer

Best practices for image generation prompts

Image generation

You can build image generation prompt bots using DALL-E-3 or StableDiffusionXL. Both bots support templating when used as a base bot.

Prompting tips

  1. Be descriptive, rather than instructive. Unlike for text generation bots, the style prompt should describe the desired image, not be an instruction to the bot. This is especially true for StableDiffusionXL. For example, to generate paintings in the style of Vincent Van Gogh:

Instead of:

You are VanGoghBot. You will generate paintings in the style of Van Gogh.

Try:

painting, Van Gogh
  1. Be specific in your prompt: The more specific and detailed your prompt is, the better the chances of getting the desired image. Instead of using a generic term like landscape, try specifying elements like sunset over a beach with palm trees. You can also specify multiple elements or style descriptions separated by punctuations, e.g. beautiful sunset, rain, painting, Van Gogh.
  2. Utilize the negative prompt (StableDiffusionXL only): If there are certain elements you don’t want in your image, use the negative prompt feature. For example, if you don’t want any buildings in your landscape, include buildings in your negative prompt.

How do style prompts and user prompts get merged?

By default, style prompts and user prompts get concatenated.

For example, with a style prompt:

cartoon --no color

and a user prompt:

dog --no cat

The final prompt will be:

dog, cartoon --no cat, color

Note: the --no parameter is only accepted by StableDiffusionXL.

Templating

Poe also supports Jinja templating for image generation bots to provide bot creators with flexibility over how style prompts and user prompts get merged.

Specifically, specify a {{user_prompt}} in either (or both!) of the style prompt and negative prompts to indicate where the user’s prompt and negative prompts should go, respectively.

For example, with a style prompt:

happy, smiling {{user_prompt}} on a skateboard --no scooter, {{user_prompt}}, cartoon

and a user prompt:

dog --no cat

the final prompt will be:

happy, smiling dog on a skateboard --no scooter, cat, cartoon

Negative prompts (StableDiffusionXL only)

Specify a --no parameter to indicate elements that should be avoided in the generated image. For example:

Updated about 1 month ago

Leave a comment