Cool images, and nice explanations, thanks for posting!<p>Couple of questions:<p>1. Why do you have to start with a small initial image? What happens if you start optimizing a large image?<p>2. Intermediate layers typically have lots of feature maps (64-512). Each individual feature map can be thought of as a color channel. In your experiments, you always optimize triplets of feature maps (I assume you use the same code for inputs to every layer). So this is not exactly how it's happening during normal training - there's not easy way to visualize an image which has 512 "color" channels, right? Other than looking at each individual feature map (where there would be no colors).