In case video preview does not show on the model card, here's a gif showing the left panel which displays the original frame, the center panel displaying the same frame downscaled to the input resolution of the discrete autoencoder, and the right panel showing the output of the autoencoder (what the agent actually sees):
