His stated purpose was image compression (although I didn’t see evidence that it worked). If the distribution encoded by the is smaller than your image, then you can send the small set of model parameters (instead of the image) and then use the model to reconstruct the target image.