Unsupervised Generated Image Editing Method Based on Multi-Scale Hierarchical Disentanglement

Abstract

In order to solve the problem of semantic entanglement in generated image latent space of the StyleGAN2 network, we propose an unsupervised generated image editing method based on a multi-scale hierarchical disentanglement network structure. In our method, we first combine the mapping layer with the style mapping layer of each resolution branch in the StyleGAN2 network, and utilize the weight matrix eigen decomposition method at each scale independently to achieve the first-level disentanglement of image attributes and obtain the semantic direction vector of the scale. Then, we use Schmidt orthogonal decomposition based on the adjacent scale eigen vector to achieve the second-level disentanglement of image attributes. The result show that, compared with other mainstream unsupervised image editing methods, our method can achieve precise image editing at multiple scales, and the measurement of disentanglement between each attribute has also reached the best.