The document discusses a probabilistic U-Net model for segmenting ambiguous images. It introduces a probabilistic U-Net that combines a U-Net with a variational autoencoder (VAE) to model segmentation as a probabilistic inference task. The probabilistic U-Net uses the U-Net to encode an input image into a latent space and decode a segmentation map, while the VAE models the latent space as a probability distribution to account for ambiguity. It is evaluated on the CityScapes dataset for segmenting lung abnormalities and other medical images.