Thank you @ak92501 for sharing our new work!
A summary of our findings:
We learn "parts" and "appearance" factors through a semi-nonnegative tensor decomposition of a pre-trained generator's feature maps. These factors can be used for a number of tasks:
🧵[1/5]