Abstract: Facial image translation aims to process the input facial image through a series of conditional operations to obtain the desired target facial image. However, existing methods often face challenges such as insufficient model generalization and attribute coupling. Based on this, a method for attribute controlled facial image translation based on latent space is proposed. Firstly, feature vectors are obtained through a feature pyramid encoding network to form a latent space; Secondly, based on the feature representation ability of the latent space, the feature vector is classified and learned, and the attribute normal vector is obtained to realize face attribute control. Afterwards, two steps, attribute normal vector decoupling and retraining, are used to solve the problem of facial attribute coupling. This method achieves fine control of facial attributes while improving the quality of image translation, and verifies its generalization in the task of sketching to real faces. Through comparative experiments and analysis of mainstream facial image translation methods such as AttGAN, the results show that this method improves the evaluation indicators such as Fréchet Inception Distance(FID) by 2% to 50% compared to existing methods, and improves the accuracy of attribute generation by 3% to 30%. This proves that this method effectively improves the performance of facial image translation under attribute control.