Disentangling Video with Independent Prediction