Diffusion-Based OOD Detection

Diffusion-based out-of-distribution detection studies how generative models can flag unfamiliar images, medical scans, actions, and structured data before a downstream system makes unreliable predictions. Methods score diffusion likelihoods or denoising errors, track trajectories through the denoising process, or compare internal representations and reconstructions with in-distribution patterns. These approaches are evaluated on settings such as near- and far-shift image benchmarks, medical imaging, inverse problems, and offline reinforcement learning, often alongside classifier confidence and distance-based detectors. A central challenge is that high likelihood or low reconstruction error need not imply semantic familiarity, while richer trajectory or feature methods can add computational cost and still require robust calibration across shifts.