This paper introduces a new domain generalization (DG) benchmark, PACS, and a novel DG method based on low-rank parameterized CNNs. The PACS dataset includes four domains: photo, sketch, cartoon, and painting, providing a more challenging and realistic scenario for DG compared to existing benchmarks. The authors argue that current DG benchmarks are too simple and not representative of real-world scenarios where target domains are sparse and visually distinct. The proposed method, which uses a low-rank parameterized CNN, outperforms existing DG approaches on the PACS benchmark. The method is based on the idea of parameterizing CNN weights using domain-specific and domain-agnostic components, allowing for flexible knowledge sharing across domains. The low-rank approach reduces the number of parameters and provides an additional route to knowledge sharing. The paper also evaluates the effectiveness of the method on existing benchmarks, showing that it performs well even on simpler datasets. The results demonstrate that the proposed method is more effective than existing DG approaches, and the PACS benchmark provides a more challenging and realistic setting for DG research.This paper introduces a new domain generalization (DG) benchmark, PACS, and a novel DG method based on low-rank parameterized CNNs. The PACS dataset includes four domains: photo, sketch, cartoon, and painting, providing a more challenging and realistic scenario for DG compared to existing benchmarks. The authors argue that current DG benchmarks are too simple and not representative of real-world scenarios where target domains are sparse and visually distinct. The proposed method, which uses a low-rank parameterized CNN, outperforms existing DG approaches on the PACS benchmark. The method is based on the idea of parameterizing CNN weights using domain-specific and domain-agnostic components, allowing for flexible knowledge sharing across domains. The low-rank approach reduces the number of parameters and provides an additional route to knowledge sharing. The paper also evaluates the effectiveness of the method on existing benchmarks, showing that it performs well even on simpler datasets. The results demonstrate that the proposed method is more effective than existing DG approaches, and the PACS benchmark provides a more challenging and realistic setting for DG research.