So, let's talk about where SD came from and that whole "if you make over a million, you gotta pay" thing.

Someone asked for a full breakdown of Stable Diffusion’s history, and I think it’s worth laying out. It started in Germany in 2021, a collaboration between a bunch of companies and universities, with the core idea being latent diffusion models. Runway and EleutherAI were both involved. Stability AI provided compute early on, then hired the original researchers, and now they’re the official maintainers.

The model uses what’s called an open license, with some usage restrictions under OpenRAIL. One thing a lot of people miss: starting from version 3.5, companies with annual revenue over a million dollars have to pay Stability.

This openness spawned a whole ecosystem—fine-tuned versions, upscalers, ControlNet, LoRA—all grown by the community. For me, understanding a model’s license and commercial boundaries is just as important as seeing how it performs, otherwise you’ll mess up and not even know where you went wrong.

A lot of people really overlooked the “million in revenue” paid tier.

I still don’t really get how strict the OpenRAIL restrictions actually are.

The community ecosystem is what’s kept SD alive all this time.

If you don’t read the license carefully, using it commercially can easily get you into trouble.

Honestly, that whole latent diffusion approach is a total game-changer.

Stability going from providing compute power to taking over — now that’s an interesting plot twist.