Double InfoGAN for Contrastive Analysis

TL;DR

This paper introduces Double InfoGAN, the first GAN-based method for Contrastive Analysis (CA) that jointly learns common latent factors $z$ and target-specific salient factors $s$ to distinguish shared versus distinctive content between two domains. By integrating GAN adversarial training with InfoGAN-style mutual-information regularization and an auxiliary encoder, the approach achieves higher-quality image synthesis and more accurate factor separation than state-of-the-art CA-VAEs, demonstrated across CelebA, CIFAR-10–MNIST, Brats, and dSprites–MNIST datasets. The method enables robust latent-factor separation, domain swapping, and healthy/target-counterpart generation, with practical relevance to medical imaging and related fields, while acknowledging identifiability limitations and proposing future work toward diffusion models and information-theoretic identifiability. Datasets and code are made available online to support reproducibility and further research.

Abstract

Contrastive Analysis (CA) deals with the discovery of what is common and what is distinctive of a target domain compared to a background one. This is of great interest in many applications, such as medical imaging. Current state-of-the-art (SOTA) methods are latent variable models based on VAE (CA-VAEs). However, they all either ignore important constraints or they don't enforce fundamental assumptions. This may lead to sub-optimal solutions where distinctive factors are mistaken for common ones (or viceversa). Furthermore, the generated images have a rather poor quality, typical of VAEs, decreasing their interpretability and usefulness. Here, we propose Double InfoGAN, the first GAN based method for CA that leverages the high-quality synthesis of GAN and the separation power of InfoGAN. Experimental results on four visual datasets, from simple synthetic examples to complex medical images, show that the proposed method outperforms SOTA CA-VAEs in terms of latent separation and image quality. Datasets and code are available online.

Figures (9)

• Figure 1: Two examples of datasets for Contrastive Analysis. First figure: Brats dataset menze2014multimodal. Top: MRI images of healthy brains (control dataset). Bottom: MRI images of brains with tumor (target dataset). Second figure: CelebA dataset. Top: control dataset with regular faces (no smile, no glasses). Bottom: target dataset that contains smiling faces with glasses.
• Figure 2: Double InfoGAN. Our model takes two inputs: $\mathbf{z}$ (common factors) and $\mathbf{s}$ (salient factors). The generator $G$ produces fake images that, together with the real images, are passed to a discriminator and encoder. The discriminator has two modules: $D$ for detecting real from fake images, and $C$ for classyfing images in the correct domain (i.e., $X$ or $Y$). The encoder $Q$ has two modules, $Q_z$ and $Q_s$, to reconstruct the latent factors $(\mathbf{z},\mathbf{s})$. $D$, $C$ and $Q$ share all layers but the last one.
• Figure 3: Image reconstruction and swap with the CelebA with accessories dataset.
• Figure 4: Fake images generated by our model. In each row, we use the same common feature $\mathbf{z}$ for all images, $\mathbf{s}=0$ for $X$ and different salient features $\mathbf{s} \neq 0$ for $Y$.
• Figure 5: Image reconstruction and swap with Cifar-10-MNIST.