Prompting: A not so important discussion

When you are using AI to generate your own prompt, you are off-loading cognitive abilities to the AI. And while this is not terrible as sometimes we can grasp a concept or we can be bothered and let the AI give us a hand, doing this in a regular basis makes you even more lazy to actually read the prompt and fix mistakes. It is much easier to just feed back the prompt again into the AI and ask it to fix it for you.

Now is important to understand that, if you are lacking info on how to improve your prompt is totally legit to use the AI to help you obtain more keywords. One aspect that a lot of people including myself get swept by the prompting is making an interesting scene, the small details, we let the model handle them, but in theory we should be the ones handling that.

The Skillset

On principle, prompting is a mental skill. Is not one that is absurdly hard, but is a skill that requires you to know:

• How words interact inside the model
• An understanding of how the model will treat your keywords
• Knowing and remembering specific keywords that will make your model do Y or X
• Cheats or tricks to make the model actually behave like you want (this imply a use of weights to fuse concepts for example)
• Getting familiar with your model, what you can accomplish and what you can't
• Controlling weight, controlling the order of your keywords, deciding what aspects of a subjects are more important
• Know which model is best for your use case

Keywords and Tags

I prefer the term Keyword because it mixes better the two styles of prompting. A Keyword is in fact a word with Weight. A good example of Keywords with Weight are Skull, ocean, sky. These words alone always generate even if you didn't do well in a natural language prompt, mostly due to the fact that they have a strong weight.

The Gap and Sanitization

There are a lot of Loras and Models that sanitize your prompt in a way that helps you get a good result regardless of how unorganized or strange your prompt may be. Models that have been trained on high quality images have the disposition of making such high quality images. If you don't trust me, do this exercise: make an empty prompt with only quality tags and negatives and a different seed each time. You will understand what I mean.

Being honest here: I prefer some of the ugliness. Is great to see a good image, but I love the imperfection in the art, it makes it more human despite being made by an AI. Maybe is just me being me, but I really like the less polished and more experimental look, it brings me closer to the artist that I like.

Anatomy of a Prompt

Each person makes prompt their own way and there isn't a perfect way of making prompts. Like botmaking, this is pretty much trial and error until you get something you are comfortable working with. Some techniques that work okay:

• Dividing into sections: Subject, background, quality tags, composition. AI grasp better the concept of keywords that appear early.
• Top to bottom: Describing subjects from top to bottom. From my experience this makes bleeding less egregious.
• Seed testing: Refine your prompt in the same seed. This helps a lot at actually understanding what your change made in the final image.

Learning Exercises

• Pinterest Reference: Use any image as reference to make a prompt. Don't put it into an AI first, do it yourself, then you can see how off you were from a machine.
• Interpretation: Study other images or prompts in order to replicate them with prompting. Don't copy their prompts, interpretate the image yourself.
• The Hard Mode: Use an artistic prompt (a concept) and decide how to represent it. A character design? Or the image itself represent the concept?

Final Thoughts

My intention is not to say that prompting is a "hard job." Prompting is not Image Generation as a whole, but it is an important part of it. My workflow is not that dependent on prompting since I draw most of it and then use AI to polish, but prompting remains a fundamental aspect of Image Gen that can't be overlooked or half-assed.