In recent years, numerous methods for deepfake detection have emerged, but they often face performance issues when evaluated across different datasets. This paper introduces two components based on the frequency and RGB domains that achieve excellent generalization. Specifically, a Self-Blended Generator Based on Discrete Wavelet Transform is proposed to generate Deepfakes from a frequency perspective, providing indistinguishable training images that enhance the discriminator’s generalization ability. Additionally, a Swin-SRM Dual-Stream Discriminator is presented, which performs multi-stage cross-attention fusion between the RGB and frequency domains. Comprehensive experiments demonstrate that the two components introduced achieve more superior performance than existing models, particularly in terms of robustness and generalization across different datasets.