GAN-based Transfer of Interpretable Directions for Disentangled Image Editing in Text-to-Image Diffusion Models

The rapid advancement in image generation models has predominantly been driven by diffusion models, which have demonstrated unparalleled success in generating high-fidelity, diverse images from textual prompts. However, these models are often characterized as black boxes due to their complex, less-understood mechanisms, highlighting a significant gap in interpretability research. In contrast, Generative Adversarial Networks (GANs) are praised for their well-structured latent spaces that offer rich semantics, enabling more straightforward exploration and understanding of model behaviors. GAN2Diff bridges this gap by transferring the structured, interpretable latent directions from pre-trained GAN models—representative of specific, controllable attributes—into diffusion models. This approach enhances the interpretability of diffusion models, preserving their generative quality while providing new avenues for exploring and manipulating complex image attributes.