- We start with the same setup such as Generator feeding into VGG19.
- The style loss will be defined by gram (if we follow online doodle)
- The input image is noise + semantic map a. map is divided into CxHxW with C represending the number of distinct region.
- For training we'd need to generate random maps (using diamond square -> KMeans clustering)
What's next?
- Take a heat map -> n color cluster -> separate into MxHxW tensor