For large language models you have a good point. The space is dominated by closed source company OpenAI, the open source ai models don't come close. This is indeed a worrying development. The current models are simply really really expensive to run, so hobbyists can't contribute in a meaningful way.

But for image generation you basically only have stable diffusion and midjourney. And I'd argue stable diffusion is much more widely used due to the control it gives and it can easily be run on consumer hardware. Customizing a model is also possible and takes only a few hours on a modern gaming computer.