i'm the anon that was talking earlier about training an unslop kontext lora
i think it has potential but my slop vs unslop dataset isn't good enough at the moment, what i tried was just running unslopped images through 0.35 denoise and prompting flux for plastic skin, oversaturated, bokeh, blur etc, then training on the reversed dataset
this is the result after 1500 steps
one of the problems is my slopping method doesn't introduce enough blur so it doesn't learn to unblur
if any anons have any slopping method while retaining the original image composition without too high denoise that'd be helpful
picrel is the first test lora after 1500 steps on only 30 images