Researchers at NVidia are showing off Perfusion, a text-to-image model they say is 100KB in size and takes four minutes to train. The model specializes in customizing a photo. For example ...