The new version of the image generation is so incredibly garbage it’s genuinely unusable. Instances of AI artifacting have multiplied exponentially, the system barely follows prompts, characters are jumbled messes of blurry features, it’s awful. If it’s an issue of expense, I would be willing to donate along with countless others if you took a collection. Please as a community we beg of you to change it back. I would be happy to pay for access to the better model, or a subscription for improved functionality, just anything but this.
I don’t think it’s bad. I think the AI’s results vary based on prompts. If you’re extra detailed in your prompts then you will get good results. Otherwise, if you aren’t too hyper specific, then you’re going to get a bunch of junk images. The whole point of the AI image generator is to write good prompts.
Basically, good prompts = good results, bad prompts = bad results.
You don’t have to write an entire paragraph. All you need to do is write what you want the AI to generate, preferably either in a list format like this:
Caucasian female, brown hair, green eyes, wearing a blue dress, sitting on a bench at a park, holding a parasol
Or as a character description, like this:
A Caucasian female with brown hair and green eyes wearing a blue dress sitting on a bench at a park and holding a parasol
Of course, results vary as the AI’s outputs tend to be randomized, but as long as you write good prompts, you won’t get bad images. Hope this helped.
Klein: the current model is alao a super distillation model, that’s relatively new under the hood it is a highly technical model at that, and so lota solutions experimenting, and developmental research is needed. Although if you are really good at summary prompting and happy with its more pre deterministic blends of base style locks Its distillation makes those powerful and they pick up on crappy prompting even and make good of it. I’d just keep in mind the new model is still in it’s early days, somewhat still being trained, and even its developmental teams are still studying the potential and upper limitations of the model designing varies sub/routines for it. It being a highly technical model also means you may find there’s actually more features to leverage under the new models hood, you just need to know how too either or JSON/Format/or generally prompt your way into those features.
admittedly even I got frustrated at it earlier today,… Just needed a break from image gen and technical rebuttal in general I guess. Oh yeah LOL “yes I’am back on lemmy world” Goos break It was haha .
“relatively new under the hood it is a highly technical model at that, and so lota solutions experimenting, and developmental research is needed.”
In other words, they are using the production environment as a test platform. Brilliant.