qwen image uncensored

Lustify SDXL has a trashy dataset, it does seem so indeed. And the CLIP of those older architectures is an awful bother to wrangle. Booru tag clouds don’t really cover artistic expression as well as one would hope. Creativity needs complex notions expressed. And it’s so badly conflated between their associations… All-in-all, SDXL has always been a mess of an architecture, even worse than SD1.5, and people attempts to enrich it with finetunes were largely futile. Very difficult to get a quality image out of it, except by sheer luck.

Qwen Image I haven’t tried outside of Venice and here it’s currently very dubiously provided with locked parameters, so I can only judge it on this weird parameter set. It’s got a much more modern natural language processor, which is very welcome. But the dataset has some deep holes in it, no pun intended. What it does know — it depicts well but you have to insist so hard for even that..! And then it has super unhealthy beauty standards: any woman is either a stick figure or a walrus — it just splits people into only those two categories existing. And, for example, to very slightly increase thigh volume of a beautiful subject, where the model already decided this isn’t shaping up to be a walrus, I’m having to pile grotesquely overexaggerated synonyms like “gigantic extremely huge humongous colossal monstrously thick curvy voluptuous thighs” — you’d be surprised how aesthetically pleasing and close to the typical average healthy human norm the result is. It’s a mess to have to jump those hoops. Hopefully the parameters restriction is lifted promptly — incredibly dubious to be limited in that regard even on a PRO plan — and then I can try just asking for more compliance with the adherence slider.

FLUX I’m not so hot on, personally. However, I haven’t messed with its finetunes — maybe they’re better than SDXL finetunes, but the base model didn’t seem to be. Main issue for me is that it’s stuck with the ancient CLIP architecture like SDs that can’t fully grasp natural language still. I hear Chroma finetune is promising.

Sadly, I’m not aware of any good uncensored NLP-architecture model. Looking at Venice to find it for us. I can imagine strong Qwen Image finetunes being fire…

P. S.

Just took a look. Qwen finetunes exist but don’t seem ready for showtime. However, there is a Multi Concept NSFW LoRA that does seem to fill in a lot of gaps in the base training. Venice could plug that into their pipeline.