is there a difference between the images the model was trained on and the images its getting in production?