Visualizing the Fréchet Inception Distance. We evaluate Generative AI not by comparing pixels, but by extracting features using an Inception Network and comparing their statistical distributions in a latent space.
Euclidean distance between centers $||\mu_r - \mu_g||^2$. Penalizes AI for generating the wrong *kind* of features.
Difference in distribution shapes (Trace term). Penalizes AI for lacking diversity or having the wrong variation.
We don't calculate this on raw pixels. We pass images through a pre-trained image classifier called InceptionV3. We chop off the final classification layer and look at the 2,048-dimensional feature vector it produces. This visualizer is a 2D simplification of that 2,048D space.