Domain generalization-based face anti-spoofing has attracted the attention of researchers. Traditional domain generalization methods adopt adversarial training to explore domain-invariant feature spaces, but inevitably disrupt the semantic structure of the model. Instead of directly editing visual features, we propose a Semantic Guided-CLIP (SG-CLIP) based on the Contrastive Language-Image Pre-Training (CLIP) model. SG-CLIP uses language-guided classifier weights to re-calibrate visual features and boost generalization performance. Specifically, in the text branch, SG-CLIP uses content text prompts to describe the content information of each input sample, enhancing the model’s understanding of content attributes. In the visual branch, C-Adapter is used to transfers the CLIP model to face anti-spoofing with a minimal set of trainable parameters, enabling the image encoder to perform fine-grained analysis of subtle facial differences. Experiments on multiple datasets show that the effect of SG-CLIP is better than the current advanced algorithms.