"Semantic Image Synthesis with Spatially-Adaptive Normalization" unofficial paper implementation with Fastai v2.
- [] Train with perceptual loss.
- [] Add feature matching loss. Did not contribute much to scores according to the original paper.
- [] Add image encoder.
- [] Try to modify the (multiscale) discriminator to make something like what was done in "A U-Net Based Discriminator for Generative Adversarial Networks".