Precompute activations

Based on Lesson 2 i have noticed the following when i compare approach 1 vs 2

approach 1: do augmentation, set learn.precompute=False and fit

approch 2: do augmentation, delete tmp directory, keep precompute=True, and fit

Approach 2 is not only way faster but also gives much better results. WHY? I am confused so any help would be appreciated