Search papers, labs, and topics across Lattice.
This paper introduces OSD-IRF, a one-step diffusion model for unsupervised industrial anomaly detection (uIAD) that leverages inverse residual fields (IRF) to accelerate inference. The method trains a DDPM on normal data, then predicts the IRF of test samples based on the DDPM's noise estimate, performing anomaly detection by evaluating the probability density of the IRF. The key finding is that anomalies are more distinguishable in the IRF space, enabling a single-step diffusion process that achieves state-of-the-art or competitive performance with a 2x speedup compared to iterative diffusion approaches.
Anomalies become surprisingly distinguishable in the inverse residual field of a diffusion model, enabling a 2x speedup in unsupervised industrial anomaly detection.
Diffusion models have achieved outstanding performance in unsupervised industrial anomaly detection (uIAD) by learning a manifold of normal data under the common assumption that off-manifold anomalies are harder to generate, resulting in larger reconstruction errors in data space or lower probability densities in the tractable latent space. However, their iterative denoising and noising nature leads to slow inference. In this paper, we propose OSD-IRF, a novel one-step diffusion with inverse residual fields, to address this limitation for uIAD task. We first train a deep diffusion probabilistic model (DDPM) on normal data without any conditioning. Then, for a test sample, we predict its inverse residual fields (IRF) based on the noise estimated by the well-trained parametric noise function of the DDPM. Finally, uIAD is performed by evaluating the probability density of the IRF under a Gaussian distribution and comparing it with a threshold. Our key observation is that anomalies become distinguishable in this IRF space, a finding that has seldom been reported in prior works. Moreover, OSD-IRF requires only single step diffusion for uIAD, thanks to the property that IRF holds for any neighboring time step in the denoising process. Extensive experiments on three widely used uIAD benchmarks show that our model achieves SOTA or competitive performance across six metrics, along with roughly a 2X inference speedup without distillation.
