EchoChange: Diffusion Language Model for Factual Remote Sensing Disaster Captioning

A recent paper published on arXiv (2608.01856v1) presents EchoChange, a multimodal discrete diffusion language model aimed at bi-temporal remote-sensing disaster change captioning. This model tackles the difficulty of detecting sparse and spatially confined changes in extensive pre- and post-event imagery, converting them into clear and factual narratives. Traditional change captioning techniques often rely on autoregressive decoding, where an initial misjudgment of a changed object or event leads to irreversible inaccuracies in subsequent text, compounding visual confusion into a series of factual mistakes. In contrast, EchoChange approaches change captioning through iterative masked-token denoising, continuously refining the entire caption based on the image pair, enabling it to reassess uncertain information and rectify flawed intermediate predictions.