For all the press releases the images displayed are always very impressive, however whenever I try a similar (at least from my point of view) prompt I get undefined blobs full of mistakes. Is my prompt-fu that low or is that a more widespread issue?
Even the press release images have multiple pretty significant issues. They look good at first glance but are pretty much nonsensical if you take a closer look.
Why is the spoon writing on the back of a clipboard, for example?
Especially for previous generation of models (from Dall-E 2 to SD 2.1) without any sort of finetune, my experience was that you can get good results only and only if you have an amazing prompt. This obscure problem seems to be slowly disappearing with newer generation (SDXL) models or the existing fine tunes (RealisticVision).
It's a widespread issue. With DALL-E 2, yes, the vast majority of the time you do get undefined blobs full of mistakes. You can get some amazing things, but only with clever prompt engineering.
This is about DALL-E 3, which is just announced. Nobody's played with it yet so we don't know if it's a lot better or not.
Prompt-fu is definitely part of it. It takes dozens to hundreds of hours of playing around before you have a good eye for it... which isn't going to happen with something you have to pay for per image, let's be honest.
