<p>Domain generalization in person re-identification (DG-ReID) aims to transfer models trained on labeled source domains to unlabeled target domains, addressing significant domain discrepancies across datasets. However, directly applying an effective single-source domain generalization model to multi-source tasks often leads to low stability and reduced generalization performance. To enhance the performance of multi-source domain generalization and reduce computational costs, this paper proposes a Dual-Branch Structural Asymmetric Mutual Learning (SAML) model to learn generalizable feature representations for unseen target domains. This marks the first application of a structurally diverse mutual learning framework within the multi-source DG-ReID field, featuring two depth-structure-diverse networks that independently learn from multiple sources while being mutually constrained. A Scale-Adaptive Matcher based on transformers aligns these networks without interference from varying feature sizes and supports supervised learning via KL-divergence loss. Extensive experiments demonstrate that the SAML model significantly outperforms state-of-the-art methods on benchmark datasets, showcasing its advanced performance and high generalizability.</p>