>>16854
>But how do you send a loli photo to Qwen without him saying he cant modified photo with cp on it?

It's about locally deployed models, not online.
Because of that it just can't "say anything". Just obey as it can 😄

Also I won't recommend you to ask any online model to do such things.

>>16855
>I did find, however, that details like dripping cum tend to get lost going from color pencil to photorealistic, but if I dare add to my prompt a request to keep the cum or put it back, I end up getting no PR conversion at all.

Somehow, when you're trying to convert image to some drawn style (line, pencil, etc), it percepts things "as is", including body fluids, reproduces them perfectly.
But if there's NO additional requests like "make face more younger" - in this case it starts to rethink mentioned areas and remove some certain details, like body fluids.

But when asked to render the image as PR, some absolutely other part of its digital brain are involved, and it is starting to "rethink" the whole content, removing fluid things and "averaging" faces too. Same about "oil painting" style.
Probably this is because of image pairs present in datasets.

That's why I'm mostly stopping on "colored pencil" style. Not also like "photoreal". Also I can do anything I want in 3D, so this is just for fun.

Also, this may happen because you're using not the right checkpoints. What set do you use (monolite one, or CLIP/UNET, safetensors or GGUF and what the names)?