Much like LLMs are still pretty bad at logic. It seems like we're plateauing in that regard...
The thing about assets is that they have to be coherent with each other. You can't just generate them one at a time in a bunch of different styles and sizes or they'll look terrible. In the simplest case you want to generate at least an entire spritesheet for a single character or object. And that's still fundamentally impossible with modern generative AIs
I really wish there was a tool where I can select the theme, which objects I need and it would spit out some half decently looking art. I even don't care if it was some simplified, cookie-cutter art as long as it was low-effort and consistent. It could be a real time saver for prototyping.
Sprite sheets are achievable or near achievable imo. SDXL was maybe good enough as a foundational model if you could leverage IPAdapter or a Lora to maintain character consistency. Although the former was never effectively tuned for styles I wanted and the latter requires some actual artist output still. Flux as a base model is far more consistent and has good enough controlnets to do sprite generation but probably still needs some a proper IPAdapter running to be stable enough for production.
Ideally it’s just a productivity tool for an actual artist
I don't understand the LLM/logic analogy, other than that there's still things about GenAI that are lacking today.
Current GenAI is not streamlined for asset generation. But I think there is a huge opportunity to go into that niche and address the problems that currently exist.
Check out the authors discussion on the subject of 'turds' [0], image artefacts created by AI that totally break immersion as soon as you properly notice them.