Reproduce “Learning Common and Salient Generative Factors Between Two Image Datasets” on OpenResearch

About this paper

Yunlong He, Gwilherm Lesné, Ziqian Liu, Michaël Soumm, Pietro Gori

This research introduces a novel framework for Contrastive Analysis (CA) that adapts state-of-the-art generative models like StyleGANs and Denoising Diffusion Implicit Models (DDIMs) to separate common and salient generative factors between two image datasets. It achieves superior image quality and disentanglement compared to previous CA methods, enabling high-fidelity reconstruction and precise, weakly-supervised semantic image editing across diverse domains, including medical imaging where it successfully distinguishes pathology from healthy tissue.

Key insights

• State-of-the-art generative models like StyleGANs and Diffusion Models can be successfully adapted for Contrastive Analysis, significantly improving image quality and factor interpretability over previous approaches.
• Effective disentanglement of common and salient generative factors is achievable with only weak supervision (dataset labels), drastically reducing the need for extensive attribute annotations.
• Sophisticated latent space regularization, including adversarial methods for statistical independence, is crucial for ensuring clean separation and meaningful semantic manipulation of factors.

Summary by alphaXiv.