What Goes Into an AI Art Style
I have spent a lot of time recently working on image art styles for AI image generation. From the outside, rendering an image in a particular art style looks like a single step. You type the name of the style into a prompt, the model denoises, an image comes back. Each of those style names is actually a carefully written prompt, and tuning the words inside it has been the thing I have spent the most time on.
These models were trained on billions of image and caption pairs scraped from the internet. Over the course of that training, the model internalizes associations that go beyond what the words in a prompt literally describe. The name of an art movement pulls in not just the visual technique, but the era it was popularized in, the subjects typically depicted in that style, and the lighting and composition that tend to accompany it. When you ask the model for "watercolor" or "ink illustration" or "3D render," you are inheriting that whole bundle whether you wanted it or not. Most of the work I have been doing on each individual style is figuring out which of those associations help and which fight.
A pattern I have run into repeatedly across the styles I have been tuning is that a style prompt that names a specific era or studio will quietly drag in the subject matter that era or studio is known for. Ask for a render in the signature visual language of a contemporary animation house, and the model fills the frame with that studio's signature contemporary subject matter, regardless of what the rest of the prompt actually described. A historical scene comes back looking historical-adjacent rather than historical, because the prompt anchored the model in a register where historical settings barely appear in the training data. The fix is rarely to add more tokens. It is usually to swap one anchor for another that pulls in the right combination of style and subject at once, which only becomes obvious after generating test images and watching how the model fails. Every style I have worked on has gone through several rounds of that.
Every style prompt is a compromise between style fidelity and content fidelity. Pushing harder toward a particular visual style makes the output look more convincingly like that style on the scenes where it works, and increases the number of cases where the model pulls the rendered image away from what the rest of the prompt described. The model takes a string of text and returns an image, and the image either looks right or it does not. The engineering that decides which of those happens is the part nobody is supposed to notice. That is also why it has gotten so much of my attention lately.