Animating Doodles with Autoencoders and Synthetic Data (2018)

The particular synthetic dataset I used consisted of grayscale images (64×64) with random lines, curves and ellipses painted on it. Here’s the same tree animation as above, with added brownian bridges using different standard deviations for the random walks:

The added noise in the interpolating paths manifests as shaky animations as opposed to noise in the actual image. The encoding in the latent space is sort of a compressed representation of the input image, with information about the curves and strokes in the image.

Source: rajatvd.github.io