This study addresses the challenges of facial detail loss and discriminator bias in image style transfer tasks. Building upon the DualStyleGAN network, the proposed the MixStyleGAN network, which improves upon the existing method by introducing a de-bias discriminator. The de-bias discriminator aims to remove style bias by generating mixed features of the original and reference images in the discriminator’s feature space, ensuring consistency between the prediction of mixed features and the original image. The AdaIN network is employed for feature map fusion across different resolution layers to achieve de-biasing effects. Additionally, localized discriminators are introduced to preserve facial details. Separate discriminators are employed for specific facial attributes such as eyes and mouth, enhancing the representation of local details through adversarial training. Experimental results on the FFHQ dataset demonstrate that MixStyleGAN achieves a 29.03% improvement in Arcface metric and a 12% improvement in LPIPS metric.