I'd like to create a keras model to produce infinite fresco based on some photos found on http://earthview.withgoogle.com I'd like to be able to generate an arbitrary number of pixel rows with smooth transition between them.
My first idea was to feed let's say 10 rows of an image to a RNN an trying to make it predict the 11th, but the model was getting stuck predicting the same row after some iterations.
I will probably try to implement an autoencoder but instead of feeding it images, try to feed it rows and then modify slightly the latent space to generate each row, is it a good idea and what are your thoughts about it ?
Thanks
