It seems that a common "pipeline" is to run the outputs of SD through GFPGAN to correct the faces. Although it can't help with the number of limbs or fingers.