I'm not sure what to make of this, it works but it's not very stable. Either the characters are off-model or they are in some unnatural pose. I'll need to do some experiments to see what kind of prompts work best with this.

Edit: It seems like both of the user submitted images used Loras specifically for making those characters, that's not a good sign.

I'll keep an eye on this, I have a feeling that all it needs is either more training data or lots more Epochs for it to become more stable and it'll work fine.

going by the negative tags on the site, the model looks to be using danbooru style tags (and straight out says its using waifu diffusion) but there are other things like tails being set in the model as "tails the fox" instead of "miles prower" or the danbooru standard "tails_(sonic)", because of this deviation from the standard danbooru tags im curious if any other tags are different from the standard danbooru tags besides character names or if there are any additional tags that arent in the standard danbooru CSV in this model.

looking through huggingface, there doesnt seem to be any CSV file specific for this model to aid in tagging, would be possible to release a CSV?

V3-Beta1 puts out more cleaner, sharper, more detailed images as well understand prompts better
i asked V3-Beta1 to put amy rose sitting on a chair and pics V3-Beta1 did just that V3-Beta4 just placed her on the floor not to mention it put out an blurry image and watered out lines so i recommend to keep using V3-Beta1 in till the dev puts out a better version that bests V3-Beta1

Wow, the amount of pictures that includes my mobiusunleashed.com logo is insane, I guess it was trained with lot of my works, heh, oh well, Tech evolution can make jobs obsolete, It's tough, but no new tale. Blacksmiths, typists, paperboys, milk mans, translators, Change is constant, resilience is key.