Played with it for a bit in DreamStudio so I could control more of the settings. So far everything it generates is "high quality", but the AI seems to lack the creativity and breadth of understanding that DALL-E 2 has. OpenAI's model is better at taking wildly differing concepts and figuring out creative ways to glue them together, even if the end result isn't perfect. Stable Diffusion is very resistant to that, and errs towards the goal of making a high quality image. If it doesn't understand the prompt, it'll pick and choose what parts of the prompt are easiest for it and generate fantastic looking results for those. Which is both good and bad.

For example, I asked it in various ways for a bison dressed as an astronaut. The results varied from just photos of astronauts, to bisons on earth, to bisons on the moon. The bison was always drawn hyper realistically, which is cool, but none of them were dressed as an astronaut. DALLE on the other hand will try all kinds of different ways that a bison might be portrayed as an astronaut. Some realistic, some more imaginative. All of them generally trying to fulfill the prompt. But many results will be crude and imperfect.

I personally find DALLE to be more satisfying to play with right now, because of that creativity. I'm not necessarily looking for the highest quality results. I just want interesting results that follow my prompt. (And no, SD's Scale knob didn't seem to help me). But there's also a place for SD's style if you just want really great looking, but generic stuff.

That said, the current version of SD was explicitly finetuned on an "aesthetically" ranked dataset. So these results aren't really surprising. I'm sure the next generations of SD will start knocking DALLE out of the park in both metrics. And, of course, massive massive props to Stability.ai for releasing this incredible work as open source. Imagine all the tinkering and evolving people are going to do on top of this work. It's going to be incredible.