Is DALL-E noticeably better than Stable Diffusion and other self-hostable options?
I don't see why I would even bother going through the sign up process for OpenAI, only to be limited by their filters. Seems like they are late to the party now.
I've run thousands of prompts with DALL-E 2, thousands with Midjourney, and probably hundreds with Stable Diffusion.
My (very qualitative) feeling is that DALL-E 2 is good with composition and realism (e.g. generating photographs — you'll still get artefacts but it's less likely to look "computer graphics-y"), and is quite forgiving (you will usually end up with an image that makes sense).
Midjourney had a recent update and can now produce beautiful images with far more detail and realism than DALL-E 2 in some cases, especially for human and animal faces, but excels more on the computer art side of things. (Midjourney now has a community showcase gallery: https://www.midjourney.com/showcase/)
Stable Diffusion is a bit less forgiving than both, in my experience. Some people are able to create stunning images, but you have to invest more time into figuring out what works best.
I'm currently looking into taking images generated with DALL-E 2, then using them as a starting point for Stable Diffusion to add detail. It works partciualrly well for cartoon-style images.
You can see it significantly cleans up the artefacts the original DALL-E 2 image had. (Note: the original DALL-E 2 image is 1024 pixels square, but Stable Diffusion generated a 512 square output.)
Midjourney's recent upgrade was largely thanks to integration with Stable Diffusion. Somehow Midjourney's images still retain a more "premium" artistic feel to them though.
Great summary - just to add my 2 cents, I find Midjourney is far better than DALL-E 2 for creating 'artistic' / visually appealing images. DALL-E 2 actually struggles sometimes to be un-realistic, whereas MidJourney has no trouble being a bit weirder / out there.
Dall.e is much better at understanding what you want. And sometimes stable diffusion feels a bit overfitted on some prompt (especially with cars).
But Dall.e is often behind in terms of image quality. They are nice looking from far, but a bit more blurry or weird than stable diffusion if you look closely.
However you can use boths together. These days I tend to use stable diffusion first, but when a prompt is not going well I copy paste it in dall.e and get what I meant much easily. And then I import the dall.e generated image in stable diffusion to work it a bit more and get something a bit better looking.
In my experience, DALL·E has generated much better images, but people have varying opinions. Stable Diffusion is more configurable, so you might be able to tinker with it to get what you want more whereas DALL·E just works pretty decently.
