Don't they? I feel like half the pictures are probably nearly identical to some picture in the training data and the other half are distorted mess...