[andybak]

You're conflating "art commonly created by these AI-powered image generators" with "art that AI is capable of".

Just because most people create things that you categorise as kitsch should not fool you into thinking that's an inherent limitation. There's almost no limit to the kind of images that can be created given the right input.

Unless you want to define kitsch as "anything that an AI image generator is capable of producing"?

[pavlov]

It’s the opposite really. I’m not claiming anything about AI image generators. I’m saying that people have quite limited tastes, and so they end up going for kitsch because it looks good at first glance.

But now that it’s so easy to generate, will it mean a shift in taste? The current crop of AI art looks the same because enough people think those limited styles are good. But in ten years, users may be a lot more demanding.

Think about 1950s pop music, before the invention of electric guitars and synthesizers and sequencers. The range of musical expression is quite limited to modern ears. But at the time, it was uncontested among ordinary listeners that “good music” had that particular sound. There was an enormous gulf between avant garde contemporary music and pop music, just like in art today.

[oersted]

You are absolutely right regarding the capabilities.

Although many image generators now seem to be fine-tuned to output impressive but formulaic art by default, to ensure a good first-impression for a broad audience.

Midjourney for example has always been great at generating very "cool" art easily, but seems to be less versatile and always tends towards that same people-pleasing style.

[andybak]

Yes - I've noticed these trends as well. I still have a fondness for the quirks and limitations of the earlier generation models Disco Diffusion and it's kin.

Sadly it's become quite hard to run some of these due to bit rot and the fact that "pinning your Python dependencies" doesn't appear to have ever been a thing for large sections of the ML dev community.