Finding Channel-wise Tensor Subspaces in GANs

Generative Adversarial Networks (GANs) can map a latent vector sampled from a random distribution to a photorealistic image. Previous methods attempt to leverage the semantic features embedded in the latent space to enable local editing but encounter challenges in accurately disentangling these features. In this paper, we analyze the redundant characteristics inherent in the deconvolution mechanism and explore the high-order redundant information of feature maps to find channel-wise tensor subspaces in GANs. We propose an architecture-agnostic unsupervised approach, where tensor subspace clustering is used for the first time to disentangle the interpretable semantic features in GANs. Each cluster is regarded as a semantic mask by calculating its mean. The relationship between the latent vector with the selected cluster is established through network gradients to identify the interpretable directions. Note that the proposed method circumvents the demand of manually drawing regions of interest and only computes gradients in the shallow layers rather than the entire network. Experimental results on three GAN-based architectures show that our method can well disentangle the semantic features in GANs and achieves more precise local editing on facial images compared to the state-of-the-art methods.