just for cerioucity i tested this , this not gonna work at all.
it just send back the input image at how it it was.
just wondering how you said prompt such words that the exact same base model not trained for it !!?
it is the same for me, unfortunately
it worked once and do not work anymore
reading through the two sets of captions, copying their prompt styles worked well.
Don't just guess guys.
this is normal for flux. For me same - once and never again. First is that promps it trained on aren't promps, those are descriptions, and unsimilar prompt may not be seen. Secondly, flux kontext has built-in BAN for every nsfw or almost nsfw prompt, so it converts it to none. And if in 100 words prompt one mentioning of nswf is, it may stop whole prompt and re-generate original image
I was getting the same thing and i found that if I add the remove-clothes-kontext.safetensors lora at 0.5 strength in addition to this lora then it will work most of the time after that, maybe about 75% or more. the quality of that other lora isn't great, but it had a high success rate, so I think I'm getting the best of both worlds with this combo. sadly it's still kind of hit or miss, re-rolling tends to help.
So in that case, why don't we just add this lora twice at 0.5 strength? instead of adding the garbage quality remove-clothes-kontext one?