3 ms·
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 undef
by vander_elst 3y ago
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?
- skepticATX 3y agoEven 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?
- treesciencebot 3y agoEspecially 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).
- kristofferR 3y agoYou're probably not using DALL·E 3: > DALL·E 3 [...] will be available to ChatGPT Plus and Enterprise customers in October via the API, and in Labs later this fall.
- vander_elst 3y agoI'm definitely not using DALL·E 3 :) but this is has been my experience for all other image generators
- Filligree 3y agoPrompt-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.
- dmazzoni 3y agoIt'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.
- ShamelessC 3y agortf headline