Hacker News Reader

What does this mean? Isn't the quality of a model determined by how easy it is to get a good picture?

Not necessarily. IMO a good model needs to follow your prompt well, and that was my problem with Stable Diffusion.

I've been trying to get a good portrait picture with "neon lights" on Stable Diffusion and it is almost impossible. Meanwhile with the new Dall-e, that was possible. The picture specially with SDXL is good, but it doesn't really have neon lights...

I tried now similar prompt on deepfloyd and managed to get there!

would be interesting if you could used deepfloyd first for image composition, then apply stable diffusion after for purely stylistic modifications

Definitely possible :) I've been doing this with new Dall-e + img2img with Stable Diffusion.

Explaining: I created a model of me, and wanted to create some good realistic portrait pictures. First I tried to create a model of me using some of the custom models already exist and the result was bad.

Then I tried SD 1.5/2.1... It was better, but couldn't really get some of the prompts make real...

Then I tried new Dall-e, saved, and inserted my face with img2img on SD and it worked much better!

Reply on news.ycombinator.com