PerceptionDLM

PerceptionDLM is a multimodal diffusion language model for region perception, developed by Peking University MSALab and ByteDance. It pairs a visual encoder with a discrete diffusion language model (LLaDA-8B-Instruct backbone) and, given an image and several region masks, generates descriptions for all regions within a single denoising process rather than one at a time as autoregressive captioners do.

Two design choices enable this: efficient parallel prompting, which packs multiple masked regions into one prompt, and structured attention masking, which isolates each region's generation stream while sharing global image context. The authors also train PerceptionDLM-Base, a general diffusion VLM that they report outperforms LLaDA-V on 15 of 16 multimodal benchmarks and stays competitive with Qwen2.5-VL and InternVL3 at similar scale.

On the new ParaDLC-Bench (a multi-region extension of DLC-Bench), the project page reports PerceptionDLM-8B at 62.4% average with 2.9 tokens per forward and 276 s, versus GAR-8B (autoregressive) at 69.5% and 479 s, and LLaDA-V-8B at 35.2% and 3241 s. This indicates a trade-off of somewhat lower accuracy than the autoregressive baseline for faster multi-region captioning. Code, model weights, training data and the benchmark were announced as released in June 2026; it is a research model, not a hosted product.