Bad-OOD: Discovering Harmful Synthetic Diffusion Outliers via Confidence Calibration

Utilizing synthetic outlier samples has shown great promise in out-of-distribution (OOD) detection. In particular, impressive results have been achieved by employing diffusion models to generate synthetic outliers in the low-density manifold. However, guiding diffusion models to generate meaningful synthetic outliers remains challenging. The synthesized samples often fall either too close to the in-distribution (ID) data (risking overlap and ambiguity) or too far (leading to visually unrealistic results). Both extremes have been shown to degrade OOD detection performance. In this work, we propose a novel OOD synthesis framework that combines a pre-trained Representation Diffusion Model (RDM) with a simple yet effective classifier calibration strategy. RDM enables global semantic embedding generation without requiring auxiliary labels or text, producing diverse yet ID-relevant outliers, thereby facilitating a more compact ID-OOD decision boundary. To ensure the utility of these samples, we calibrate a binary classifier on both ID data and synthesized OODs to assign confidence-based anomaly scores. We find that mid-confidence outliers, i.e., those balancing realism and deviation, are most informative, and using them significantly boosts detection performance. Extensive experimental results validate the superiority of our calibrated OOD sampler over several strong baselines.