>>96763947>I heard people shitting all over the new model for being a blurry mess.skill issue (just a little)
I've been playing a lot with v4 (the new model), and there are three main issues:
>No one knows how the fuck v4 interprets the promptsThe tags still work, but the new model expects you to use prose to describe the scene, then maybe use tags to specify details, then maybe use the new character-specific prompts to reinforce what each character is doing, and finally maybe use the new syntax to further reinforce those actions. So if you just grab an old v3 prompt and reuse it for v4, you’ll probably get a hot mess. Or not, sometimes it just works
>One of the best things about v3 was how well it could combine different art styles and v4 completely shits the bed More often than not, it will focus on a single artist or art style and ignore the rest. Or just do whatever it wants and just draw a butt
>v4’s dataset is bigger and includes more artists, but for some reason, a lot of them output worse looking results than in v3From what I’ve tested, the tags used alongside the artist tag heavily influence the output quality (and some can completely ruin the picture). If the quality is bad, you’ll need to debug the prompt to identify which tags are causing issues, then either remove them or adjust their weight
The core of the model itself is great, and when it works, it can create amazing stuff, getting right most of the details that v3 struggled with (like wedding rings, legible text, etc...), but you have to fight with the damn thing to get it to listen to you. When it works, it's great. When it doesn’t, it spits out a blurry mess
So far, the devs haven’t said anything about the complaints and have just left everyone to figure out the new model on their own. My guess is that they either rushed the release for some reason and it needed more time in the oven, or they just don’t know how the fuck it works either
>tl;drIt's a mixed bag, but I'm happy creating pekomama porn
