High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs
Published:
address two problems
- difficulty of generating images of high resolution
- lack of details and realisim
method
- aim: diverse result given input label map
- object-level semantic editing
- global generator and local enhancer
- two scales often enough
- 3 discriminators for different scales
- easier to train coarse to fine
- feature matching in each layer of distriminator
- instance-level semantic label map -> object boundry
- instance-wise average pooling
- how to implement?