Metric-based methods are promising approaches to domain adaptation, aiming to align the marginal distributions of different domains with similar conditional distributions. In traditional approaches, the metric function is manually designed to measure the distance across domains. Adversarial methods can be considered an automatic learning approach to the metric function. Instead of relying on the quality of the metric function, we outline a generalized framework for domain randomization, which first introduces moderate perturbations as a form of randomness and then combines the advantages of metric-based domain adaptation with domain randomization. We propose a novel approach leveraging simulation environments to generate extensive, annotated data for diverse scenarios. Our focus is on simulation-to-real transfer for semantic segmentation tasks, acknowledging CNN’s texture bias. We introduce a domain randomization technique that limits meaningful texture information and devise a mapping function for image transformation. The proposed approach emphasizes geometry consistency and style transfer, providing a practical solution for efficient simulation to real-world transfer. The experiments are conducted on the domain generalization task from GTA to Cityscapes and BDD, and evaluated on the semantic segmentation performance. Our method achieves superior results in most segmentation classes compared with the benchmark models by 5.2 to 10 in terms of mIoU. Remarkably, our generalization strategy does not require access the target domain data at training time.