Mask2Former and OneFormer now available in Hugging Face Transformers for universal image segmentation

AI Impact Summary

Mask2FormerForUniversalSegmentation and OneFormer are now exposed in Hugging Face Transformers, enabling universal image segmentation in a single framework. They support instance, semantic, and panoptic segmentation, allowing teams to standardize pipelines across tasks and datasets. OneFormer adds a text-conditioned capability and generally incurs higher latency due to the extra text encoder, while backbones like Swin Transformer or DiNAT affect compute and memory. Before production, validate latency, model size, and dataset compatibility; start with COCO-panoptic checkpoints to benchmark against existing results.