Abstract

Reconstruction-based image anomaly detection models that utilize diffusion models have shown superior performance in recent studies. However, these models suffer from substantial computational overhead and inference latency, limiting their deployment in practical applications. To overcome these limitations, we introduce LAMBDA, a lightweight diffusion-based framework that leverages memory bank-guided denoising for efficient anomaly detection. Our approach introduces two key components: (1) Memory Bank-guided denoising that leverages pre-stored normal latent representations to achieve both computational efficiency and superior reconstruction fidelity, leading to more accurate anomaly localization and (2) Guidance Scale Map, which dynamically adjusts guidance intensity based on dataset distributions. Experimental results show that LAMBDA achieves an 80.8 % reduction in GPU memory consumption during training and 53.2 % during inference, while achieving 2.3

 and 3.5

 speedup in training and inference times, respectively, compared to the baseline. Despite these efficiency gains, LAMBDA maintains competitive performance, achieving accuracy on par with state-of-the-art models. These results demonstrate the potential of our approach for real-world industrial applications, offering a resource-efficient solution for diffusion-based image anomaly detection.