This paper addresses the challenge of unsupervised anomaly detection (AD) with noisy data, a common issue in real-world applications. Traditional AD methods often rely on clean training data, making them less effective in the presence of noise. To tackle this, the authors propose SoftPatch, a memory-based unsupervised AD method that efficiently denoises data at the patch level. SoftPatch uses noise discriminators to generate outlier scores for patch-level noise elimination before constructing a coreset. These scores are stored in a memory bank to soften the anomaly detection boundary. Compared to existing methods, SoftPatch maintains strong modeling ability for normal data and alleviates overconfidence in coreset construction. Extensive experiments on various noise scenarios demonstrate that SoftPatch outperforms state-of-the-art AD methods on the MVtecAD and BTAD benchmarks, showing comparable performance to these methods in noise-free settings. The main contributions of the paper include the introduction of a patch-level denoising strategy and the use of soft reweighting to improve noise robustness.This paper addresses the challenge of unsupervised anomaly detection (AD) with noisy data, a common issue in real-world applications. Traditional AD methods often rely on clean training data, making them less effective in the presence of noise. To tackle this, the authors propose SoftPatch, a memory-based unsupervised AD method that efficiently denoises data at the patch level. SoftPatch uses noise discriminators to generate outlier scores for patch-level noise elimination before constructing a coreset. These scores are stored in a memory bank to soften the anomaly detection boundary. Compared to existing methods, SoftPatch maintains strong modeling ability for normal data and alleviates overconfidence in coreset construction. Extensive experiments on various noise scenarios demonstrate that SoftPatch outperforms state-of-the-art AD methods on the MVtecAD and BTAD benchmarks, showing comparable performance to these methods in noise-free settings. The main contributions of the paper include the introduction of a patch-level denoising strategy and the use of soft reweighting to improve noise robustness.