Nicholas Guttenberg at Crosslabs:
[Stable Diffusion] has an interesting limitation that seems to have mostly gone unnoticed. If you try to ask it to generate images that should be particularly dark or light, it almost always generates images whose average value is relatively close to 0.5 (with an entirely black image being 0, and an entirely white image being 1).
The rest of the post is fascinating but quite technical and hard to excerpt. In the end, the results he was able to achieve are striking.
It is remarkable to me how quickly and completely Stable Diffusion was able to take over closed source, commercial alternatives. In my circles, at least, I almost never hear anyone talk about DALL-E 2 anymore. Maybe, in hindsight, it was inevitable that any individual company just couldn’t match the velocity of the entire open source community. It is a success story that I hope to see reproduced with language models sometime soon.