<p>Unsupervised person re-identification has received much attention due to the lack of annotated data. While pseudo-labeling relies heavily on clustering results and lacks the guidance of real labels, noisy labels limit the performance of the model. Existing studies have two problems: (1) the pseudo-labels obtained from clustering contain noisy labels, which affects the feature matching effect; (2) the samples with discriminative properties are not fully utilized. To solve these problems, this paper proposes a diffusion sample de-entanglement method (DSD), which consists of double branching sample entanglement (DBSE) and contrast diffusion de-entanglement (CDD). Specifically, DBSE designs upper and lower branch sample entanglement. The upper branch assigns pseudo-labels to new samples by inter-sample entanglement distance, and the lower branch acquires entangled samples and assigns pseudo-labels to them by entanglement degree. The goal of two-branch sample entanglement is to reduce the gap between pseudo-labels and real labels. CDD learns hard-to-discriminate samples through difficult additivity loss, thus mitigating the unreliability caused by noisy labels and improving the accuracy and robustness of the model. Finally, this paper conducts extensive experiments on several datasets to verify the effectiveness of the proposed method.</p>