This paper proposes a novel Test-Time Domain Generalization (TTDG) framework for Face Anti-Spoofing (FAS) to enhance the generalizability of FAS models beyond evaluation. The TTDG framework leverages testing data to project unseen samples to the known source domain space, improving the model's performance on unseen domains without requiring model updates at test time. The framework consists of two key components: Test-Time Style Projection (TTSP) and Diverse Style Shifts Simulation (DSSS). TTSP dynamically projects unseen samples to the known source space based on similarity to style bases, while DSSS synthesizes diverse style shifts via learnable style bases in a hyperspherical feature space. The framework is model-agnostic and can be seamlessly integrated into CNN and ViT backbones. Comprehensive experiments on widely used cross-domain FAS benchmarks demonstrate that the proposed TTDG achieves state-of-the-art performance and effectiveness. The method outperforms existing domain generalization and test-time adaptation approaches by effectively utilizing testing data to enhance generalizability. The TTDG framework is particularly effective in scenarios with limited source domains and demonstrates robustness across various domain shifts. The method introduces two new losses: a style diversity loss to encourage orthogonality among style bases and a content consistency loss to ensure projected features align with their corresponding content features. The framework is evaluated on four public FAS datasets, showing significant improvements in performance metrics such as HTER and AUC. The results demonstrate that TTDG is effective in enhancing the generalizability of FAS models without requiring domain labels or additional training data. The method is flexible and can be applied to various backbones, making it a promising approach for real-world FAS applications.This paper proposes a novel Test-Time Domain Generalization (TTDG) framework for Face Anti-Spoofing (FAS) to enhance the generalizability of FAS models beyond evaluation. The TTDG framework leverages testing data to project unseen samples to the known source domain space, improving the model's performance on unseen domains without requiring model updates at test time. The framework consists of two key components: Test-Time Style Projection (TTSP) and Diverse Style Shifts Simulation (DSSS). TTSP dynamically projects unseen samples to the known source space based on similarity to style bases, while DSSS synthesizes diverse style shifts via learnable style bases in a hyperspherical feature space. The framework is model-agnostic and can be seamlessly integrated into CNN and ViT backbones. Comprehensive experiments on widely used cross-domain FAS benchmarks demonstrate that the proposed TTDG achieves state-of-the-art performance and effectiveness. The method outperforms existing domain generalization and test-time adaptation approaches by effectively utilizing testing data to enhance generalizability. The TTDG framework is particularly effective in scenarios with limited source domains and demonstrates robustness across various domain shifts. The method introduces two new losses: a style diversity loss to encourage orthogonality among style bases and a content consistency loss to ensure projected features align with their corresponding content features. The framework is evaluated on four public FAS datasets, showing significant improvements in performance metrics such as HTER and AUC. The results demonstrate that TTDG is effective in enhancing the generalizability of FAS models without requiring domain labels or additional training data. The method is flexible and can be applied to various backbones, making it a promising approach for real-world FAS applications.