With all due respect, I've been using it for over a week and I don't think you've given it a fair shot.
There's plenty of cases it's worse than Dall-E and there's plenty of cases where it's better. Overall it seems to show less semantic understanding but it handles many stylistic suggestions much better. It's definitely in the right ballpark.
In fact I'm still using a wide range of models - many of which aren't regarded as "state of the art" any more - but they have qualities that are unique and often desireable.
Agreed. I still primarily use vqgan + clip, which is nowhere near state of the art, but produces really interesting results. I’ve spent a long time learning to get the best out of it, and while the results aren’t very coherent, it’s great at colour, texture, materials and lighting.
The last one was particularly crappy. It gave me a red house with a driveway, but no car. And the house wasn't even really a house. It superficially looked like one but was actually two garages put together.
Here's some random prompts I've had nice results from:
iridescent metal retro robot made out of simple geometric shapes. tilt shift photography. award winning
Scene in a creepy graveyard from Samurai Jack by Genndy Tartakovsky and Eyvind Earle
virus bacteria microbe by haeckel fairytale magic realism steampunk mysterious vivid colors by andy kehoe amanda clarke
etching of an anthropomorphic factory machine in the style of boris artzybasheff
origami low polygon black pug forest digital art hyper realistic
a tilt shift photo of a creepy doll Tri-X 400 TX by gerhard richter
I guess I might have spent more time reading guides on "prompt engineering" than you. ;-) I think maybe Dall-E is more forgiving of "vanilla prompts".
However I do get nice results from simpler prompts as well. I just tend to use this style of prompt more often than not.
There's plenty of cases it's worse than Dall-E and there's plenty of cases where it's better. Overall it seems to show less semantic understanding but it handles many stylistic suggestions much better. It's definitely in the right ballpark.
In fact I'm still using a wide range of models - many of which aren't regarded as "state of the art" any more - but they have qualities that are unique and often desireable.