We used to think AI would take the "easy" jobs first, like driving a truck. Driving turned out to be hard. Writing, drawing and composing turned out to be where generative AI is strongest. A widely shared post put the feeling well: I want AI to do my laundry and dishes so I can do art and writing, not the other way round.
I treat these models as novelty generators. They are useful when you lean into what they do well instead of asking them to copy what people already do. The Singapore artist Niceaunties makes work you could only make with a generative model. Designing your own image model, or combining models, is another route.
My own example is a tarot deck. I always wanted to make one and I can't illustrate. When DALL-E 3 came out in late 2023, it was the first image model with a language model in front of it. Midjourney does exactly what you type. DALL-E first passes your request through GPT, so you can ask it to consider every existing version of a card, say the Sun, and invent a new image that combines their symbols. I used the same prompt for all 22 cards of the major arcana over a couple of days. The result is symbolically correct and looks like nothing else. I could never have made it alone.
The same experiment taught me something less pleasant. Six months later I ran the identical prompt on the identical model, and the images were much worse: flatter, blander, less rich. My best explanation is cost. Each extra layer of generation costs compute, so providers tune the models down over time. Many heavy users report the same thing, and I did not believe them until I saw it side by side.
I also tried to let models write my newsletter, and stopped. I didn't like the output, but more than that I missed the writing. The writing is why I have a newsletter. So I automated what I could never do every week instead: the header images.
For organisations, the practical questions are copyright and misuse. Some services refuse anything copyrighted. Open models like FLUX and Stable Diffusion will make almost anything, and any model can be jailbroken.
Try this
Make something you couldn't make before: an image, a song, a short story. Then make it again with a different model and compare what each one added.
Since 2024
Image models now take instructions in plain language and edit existing pictures, so the gap between Midjourney and DALL-E that I describe above has mostly closed. The questions about style, consent and copyright have not.
Sources
- OpenAI, DALL-E 3 (2023)