This version also trained the text encoder (TE), which means it more accurately captures the style but it also can degrade image quality (bad limbs, greyscale, lower prompt following), you can tweak the weight of the TE with the <loraname:TE:UNET> syntax; If you are having problems with it I would recommend disabling the TE for generating with txt2img then enabling it for the upscaling process in img2img