OneFormer is a CVPR 2023 paper and model that uses a transformer architecture to perform universal image segmentation across different task types. It unifies semantic, instance, and panoptic segmentation into a single model, trained on datasets like ADE20K, Cityscapes, and COCO. The model leverages a multi-scale transformer encoder and task-conditioned training to achieve state-of-the-art results across segmentation benchmarks.