A busy week in AI. Google came out swinging, saying the company has smashed AI into everything, and OpenAI let GPT-4 lose on the world when 99% of the world either didn’t know what GPT-3 did or was comfortable with it. Welcome to the AI wars; they’re just getting started.
Whether you like it or not, generative AI is here to stay and will impact lives across the board in ways we can predict and, more importantly, can’t predict yet. Generative AI in the machine learning field of AI, and while you won’t quite be out of a job next week, keeping one eye on the horizon is your best bet to avoid being sideswiped. TBD Group has been counselling clients for a while on what generative AI can (and can’t do), and we thought a helpful and accessible 101 on prompt creation would be useful. We’ll look at image prompts here, and text prompts in a separate post.
Prompt what now? Think of it as what you put in to get what you want out. More specific prompts (‘Generate a still life drawing of an angry seagull on the right side of a modern mansion with ten windows’) will yield different results to looser prompts (‘Generate a picture of a bird on a house’). Anyone can design a prompt (there’s already even a visual prompt creator), but there’s a burgeoning bevvy of new ‘prompt engineers’ who are making bank from people’s lack of knowledge.
// How to write image prompts that are smarter than the average bear
First things first, and the only required element, what subject do you want to create? A flying racoon? A seascape? The world is your oyster.
The rest of the prompt is just extra information that will help the tool figure out what it should put in — or not put in — to your design. Whack in a comma and keep them coming for more specific designs or dial it back and you’ll create something truly unique.
- Style — an artist like (example: in the style of oil painting) or artist (example: by Henry Matisse) or a technique/movement/form like surrealism, symmetry, contemporary, or minimalism.
- Format — what type of image do you want (example: photograph)
- Booster — a focus on your desired quality, based on a platform, like Pixiv, or image quality. (i.e. trending on pixiv, high quality, masterpiece)
- Solidifier — give the tool examples of different ways of repetition (example: space whale, a whale in space)
- Vibe moods or emotions (i.e. control the soul)
- Weights — this is all about adding or taking away emphasis from different elements (example: no animals). You can also create a mix to up weight a specific element over another one, but still have some influence from both. Example: (by Pablo Picasso:30 | by Mondrian:70) — the brackets are integral
- Perspective — you can add a specific framing this way — e.g. looking up, through a window, street level view, panoramic etc.
You end up with something like this:
[format] of [subject term], a [solidifier], [perspective], by [artist:75] | by [artist:25] / in the style of [style], [variant], [booster], [exclusion]
Which looks like this in reality:
“sketch of a cat pirate, pirate cat, looking through a window, by walt disney:25 | by andy warhol:25, control the soul, trending on Pixiv — no dogs or people”
And gets you:
The key is to have fun, be creative and edit, edit, edit until you get the image you want or what was in your head to begin with.
- Use a minimum of seven words.
- Describe what should exist, not what is missing, when altering/fixing /creating prompts. Example: “cars” vs “14 cars” or “a worker” vs “a female worker”
- Keep a list of good modifiers and styles in a Google doc to save yourself time.
- Provide examples whenever possible. The quality of the results will be a lot higher.
If you’re interested in AI, you’ll love ‘What Did OpenAI Do This Week?’