📝Prompting

Various tips and tricks to get the most out of your prompts.

The Main Idea

Prompting with Retro Diffusion is very different from Stable Diffusion, Midjourney, or Dalle.

Retro Diffusion is trained primarily on natural language and does best with short, descriptive prompts.

When crafting a prompt, avoid "word soup" like "high detail beautiful image of a dog, best dog, cute dog, best quality, ultra quality, masterpiece, perfect lighting, cinematic lighting, trending on artstation." This kind of prompt is detrimental to the Retro Diffusion models and will return poor results.

Prompts that are stories, lists of tags, or lacking in detail also perform poorly.

Instead, format the prompt as a single, descriptive sentence followed by a handful of relevant tags.

For example, the previous prompt would perform better if written like this:

"A detailed portrait of a cute dog, field, grass, contrast lighting"

(Note the use of the technical term "contrast lighting." It is good to use industry standard terms instead of trying to describe the idea using common words).

Concept Closeness in AI Models

When writing a prompt for any AI model, consider the idea of concept closeness. As humans, we have our own way of relating ideas to each other. For example, if you were told to think about an apple, you would probably also think of the stem and the leaf. Similarly, if you think about a burger, you are also likely envisioning the bun and other dressings (you may even be picturing an environment related to burgers, like a restaurant).

AI models exhibit similar properties. The words you use to describe things will also contribute to various, often unexpected, aspects of the image. Being mindful of these connections will help you get better results when prompting.

Take, for example, the words "Beach" and "Shore."

These words are often used interchangeably, but in Retro Diffusion, "Beach" conjures images of tropical environments, typically with some human elements like beach chairs, umbrellas, or other elements of civilization. "Shore" typically results in forested areas with cliffs leading to water, and little to no human elements.

Using the Large Language Model

If your computer has more than 10GBs of VRAM, you have the option to use the "LLM Prompt Enhancement" in the generation menus. This will leverage a large language AI model to convert the prompt you write into a more detailed prompt formatted in a way that Retro Diffusion can better understand.

Using the "Prompt to Prompt" tool in the "Help" -> "Retro Diffusion Scripts" -> "Prompt to Prompt" menu can help you learn to adapt prompts for Retro Diffusion.