PixVL: Self-Supervised Training of Pixel-Level MLLMs via a Unified Mask--Text Consistency Cycle
PixVL proposes a self-supervised training method called PixVL to help pixel-level multimodal language models learn from unlabeled data. It introduces a unified Mask–Text Consistency Cycle, which allows the model to generate descriptions and verify its own generated masks by cycling between text and mask predictions. This enables joint learning for both region understanding and segmentation tasks.