Style is free now. Taste isn't
GPT-4o's image generation turned the internet into Studio Ghibli for a week. When every style is one prompt away, the scarce thing is judgment.
On Tuesday OpenAI added native image generation to GPT-4o in ChatGPT, and by Wednesday my feeds were entirely Studio Ghibli. Family photos, wedding photos, memes, historical photographs, pets, all redrawn in the soft watercolor style of Miyazaki’s films. Sam Altman posted that their GPUs are melting and they’re temporarily rate-limiting.
Technically, it’s a big step. The image generation is part of the same model that does text, not a separate model called with a prompt the way DALL·E 3 was in ChatGPT. It follows long, specific instructions well, renders text accurately, keeps details consistent across edits in a conversation, and can take an uploaded photo and transform it while keeping who and what’s in it. That last part is what made the Ghibli trend possible. The results look like your photo, in that style.
The ethics conversation started immediately and it’s the right one to have. Miyazaki has said strongly negative things about AI-generated animation in the past. The style is the product of decades of work by specific artists. OpenAI says it blocks imitating living artists’ styles by name but allows “broader studio styles,” which is a thin line. I wrote about consent for faces and voices before. Style is fuzzier, because artists have always influenced each other, but copying an entire studio’s look at scale, overnight, isn’t influence in the normal sense.
What I’ve been thinking about more is what this does to the value of style for everyone else.
For most of history, a distinctive visual style was scarce. It took years to develop and it was hard to copy. That scarcity was a large part of what creative professionals were paid for. This week showed that, at least for the surface of a style, the scarcity is gone. Anyone can have any look in thirty seconds. When everyone has Ghibli, Ghibli stops signaling anything.
What’s still scarce is taste: knowing which photo is worth transforming, what to put in the frame, what to leave out, when a style fits a story and when it’s a gimmick, when to stop. Almost all of the Ghibli images this week were the same idea executed a million times. The few that stood out did so because someone made a choice about content and context that nobody else made.
I wrote in 2015 that Gatys’s style transfer showed the surface of a style is measurable statistics, and that it didn’t capture why Van Gogh composed a scene the way he did. Ten years later, that’s the dividing line between what’s become free and what hasn’t. Generating is cheap now. Choosing is the job, and I think creative tools should be built to help people choose.