PerceptionDLM Just Revolutionized Region Captioning by Going Parallel — Here's Why That Matters

Current multimodal large language models (MLLMs) typically rely on autoregressive generation, which processes image regions one by one, leading to slow inference when captioning multiple areas. PerceptionDLM breaks this bottleneck by leveraging a diffusion language model (DLM) that can generate text in parallel. The model is built upon PerceptionDLM-Base, which already achieves state-of-the-art performance among open-source diffusion MLLMs. Its core innovation is a combination of efficient prompting and structured attention masking that allows the model to perceive multiple masked regions concurrently—both at the sequence level (multiple captions per image) and token level (within each caption). This design radically improves efficiency compared to sequential approaches.

To rigorously evaluate parallelism, the team created ParaDLC-Bench, an extension of DLC-Bench that includes multiple region masks per image for simultaneous captioning and speed measurement. Experiments demonstrate that PerceptionDLM maintains competitive caption quality while delivering substantial inference speedups for multi-region tasks. The authors claim this is the first work to achieve parallel region captioning using diffusion language models. Code, models, and datasets are publicly released, offering a practical foundation for efficient, real-time visual perception in applications like autonomous driving and robotics.