The model was trained to produce detailed eyes and faces. You can see visual comparisons in the images. For best results during generation, use an adapter strength/weight of 0.4–0.6.