We introduce a method that allows to automatically segment images into\nsemantically meaningful regions without human supervision. Derived regions are\nconsistent across different images and coincide with human-defined semantic\nclasses on some datasets. In cases where semantic regions might be hard for\nhuman to define and consistently label, our method is still able to find\nmeaningful and consistent semantic classes. In our work, we use pretrained\nStyleGAN2 generative model: clustering in the feature space of the generative\nmodel allows to discover semantic classes. Once classes are discovered, a\nsynthetic dataset with generated images and corresponding segmentation masks\ncan be created. After that a segmentation model is trained on the synthetic\ndataset and is able to generalize to real images. Additionally, by using CLIP\nwe are able to use prompts defined in a natural language to discover some\ndesired semantic classes. We test our method on publicly available datasets and\nshow state-of-the-art results.\n