AI art before vs after: what actually changed in 18 months?

I started messing with AI image generators in early 2022. Back then you'd type "a cat" and get a nightmare blob with three eyes and no legs. Fast forward to now. The same prompt gives me photo-realistic fur, correct anatomy, and lighting that looks like a pro studio shot. But here's what bugs me. Did the tech really get that much better, or did we just learn how to prompt it right? I see people claiming it's all model upgrades. But I also watch guys on YouTube write a single sentence and get magic. My gut says it's both. The training data got cleaned up across the board from Midjourney to Stable Diffusion. And we learned to stop asking for nonsense. Has anyone else noticed a bigger leap from learning than from new versions?

3 comments

3 Comments

dixon.iris18d ago

Prompt engineering is a bigger deal than most people give it credit for, honestly. Once I stopped trying to describe perfect scenes and just started asking for simple things like "a cat, photorealistic, natural lighting" the results got way better without any model updates. It feels like we finally figured out the secret handshake with these things (or at least I did).

nora_park18d ago

Did you just say you figured out the handshake or the secret...?

seanjackson18d ago

It depends a lot on which model you're using though.