Hi folks,
Excited to share what Soumik and I have been working on for the past few days. We present an implementation of GauGAN [1] in Keras: GauGAN for conditional image generation. Our core focus was on readability but we are happy to see it performing quite well too. We are also announcing this repository where we will publish results with bigger datasets: GitHub - soumik12345/tf2_gans: Implementations of GANs in Tensorflow 2.x.
During training, GauGAN learns to generate images that are conditioned on semantic segmentation maps and latents learned from cue images (“Reference Image” in the above GIF). What a great way to extend Variational Autoencoders isn’t it! For those who are wondering, yes, this is the architecture we used to see a few years back using which they drew some simple paintings and generated images from them [2].
Thanks to @fchollet for helping us with the reviews.
References
[1] [1903.07291] Semantic Image Synthesis with Spatially-Adaptive Normalization