Invertible Hierarchical Generative Model for Images

Supplementary material

Videos

In the below video, we show that random walks in the latent space of our model correspond to smooth changes in the decoded image. The interpolated images also remain sharp and close to the manifold of real images.

Video 1: Linear interpolation between randomly sampled points from the latent prior in our model.

The below video contrasts interpolations between real images for our model and Glow. The aliasing-artifacts of Glow are clearly visible around the midpoint of the interpolation. Our model also exhibits slightly less pixel-level crossfading between images (e.g. the shoulders in the fourth column).

Video 2: Linear interpolation between reals for our model and Glow.

The final video below complements the standard-deviation visualization of Figure 3 in the paper. Each image is generated by encoding the same image and the replacing a single layer of the latent with a sample from the prior. The samples in both models are untruncated to clearly illustrate the variance, but often result in broken images.

Video 3: Supplementary video for Figure 3. Variation by each resolution of latents.