Counterfactual generation with joint causal representation for generative adversarial networks.

Abstract

The black-box nature of Generative Adversarial Networks (GANs) hinders the controllable generation of images, particularly in semantic editing that involves making changes to multiple attributes. To address this problem, we propose a Counterfactual Generation model with Joint Causal Representation(CG_JCR). The key ideas of CG_JCR include utilizing classifier gradients as prior knowledge to learn counterfactual joint representations of semantics, using orthogonalization through continuously optimizing iterations to disentangle semantic representations, and constructing independent counterfactual representations and disentanglement modules for pre-trained GANs to implement counterfactual generation. We compare CG_JCR with its competitors in Celeba dataset using a variety of metrics and intervention experiments. Finally, we empirically validate the effectiveness of generating and disentangling joint causal representations. The code is open-source and publicly available at .