It's because the models don't have an optimal aesthetic policy. Which would be difficult, but if they did have one, it wouldn't matter how much bad input data you added during pretraining.