The supplementary video DrusenPaths.gif is part of the manuscript
"Exploring Drusen Type and Appearance using Interpretable GANs", which
has been submitted to VCBM 2024. It provides examples of the animations
that are described in that paper, illustrating the smooth variations in
visual appearance of drusen that have been discovered by our
visualization approach that is based on generative adversarial networks (GANs).

The first and third row show images G(x) sampled from the corresponding generative model G.
The generative model can be seen as a function G(x)=I taking a latent vector x and generating an image I as output.

Under each of those images is a path shown.
A path is a function f(x, d) taking a latent vector x and a distance d as input. The output is x moved along the path for the given distance d.
For a given path, it holds true that f(x, 0) = x.
A path is shown as a loop consisting of 33 images, ranging from G(f(x, -2*epsilon)) to G(f(x, 2*epsilon)).

The colored bar on top of images follows the used color bar, see Figure 3, showing the p-values calculated with the Chi-squared test.
As mentioned in the paper, this is only done in Z-Space which is why the last two paths do not have any colored bar on top.

For the depicted paths, here are the Path Model, Generative model and some parameters used to generate it:

Grow Multiple 1-3: StyleGAN trained for 2000 epochs without weighted sampling. The path model is Linear in Z space.
Grow Single 1-3: GAN trained for 100 epochs with weighted sampling. The path model is PDE in Z space.
Growth and Merging 1-2: GAN that was trained for 100 epochs with weighted sampling, and the path model is PDE in Z space.
Shift In Sizes 1-3: StyleGAN trained for 2000 epochs without weighted sampling. The path model is Linear in Z space.
Lighting 1-3: StyleGAN that was trained for 2000 epochs without weighted sampling. The path model is Warp in Z space.
Almost No Change 1: StyleGAN that was trained for 2000 epochs without weighted sampling. The path model is Warp in W space.
Abrupt Change 1: StyleGAN trained for 1000 epochs with weighted sampling. The path model is PDE in W space.