Hyperspectral anomaly detection aims to spot rare targets in complex scenes, but current models typically compress the detection signal into a single scalar reconstruction error. The authors argue this "reconstruction-as-endpoint" paradigm is fragile: scalar residuals can cause sub-pixel anomalies to vanish during spatial downsampling, and training weights can be corrupted when anomalies are not cleanly separated from the background.

To address this, the paper introduces Reconstruction-to-Vector Diffusion (R2VD), a four-stage pipeline. It starts with physical prior extraction to reduce early confirmation bias, then uses an OmniContext Autoencoder to purify residual maps while preserving small target structures. A Diffusion Transformer with a Physical Spectral Firewall models residual scores and prevents cross-spectral leakage, and a final vector dynamics stage separates targets from backgrounds by evaluating high-dimensional interference patterns rather than scalar errors.

R2VD was evaluated on eight hyperspectral datasets and reported state-of-the-art performance, with improved target detectability and background suppression. The authors have also released code for the framework.