iFlowGAN: An Invertible Flow-Based Generative Adversarial Network for Unsupervised Image-to-Image Translation

Area of Science:

• Computer Vision
• Machine Learning
• Generative Adversarial Networks

Background:

• Unsupervised image-to-image translation often relies on Generative Adversarial Networks (GANs) like CycleGAN, StarGAN, AGGAN, and CyCADA.
• These models typically require learning separate forward and backward mappings, leading to parameter redundancy and memory inefficiency.
• Existing methods assume the existence of a backward mapping B: Y → X for a forward mapping F: X → Y, which is often under-constrained.

Purpose of the Study:

• To propose iFlowGAN, a novel approach for unsupervised image-to-image translation using invertible flows and adversarial learning.
• To address the parameter redundancy and memory inefficiency issues in current GAN-based image translation models.
• To develop a method for efficiently learning inverse mappings within GAN architectures.

Main Methods:

• iFlowGAN learns an invertible flow, a sequence of invertible mappings, through adversarial learning.
• The approach utilizes concepts from contraction mappings and the Banach fixed-point theorem to derive inverse mappings.
• A Lipschitz-regularized network is introduced to enable the composition of inverses for arbitrary Lipschitz-regularized networks, saving memory for backward mapping weights.

Main Results:

• iFlowGAN achieves comparable results to original GAN implementations like CycleGAN, StarGAN, AGGAN, and CyCADA.
• The proposed method significantly reduces model parameters, saving approximately half the parameters compared to traditional approaches.
• iFlowGAN successfully integrates with existing GAN architectures without architectural modifications.

Conclusions:

• iFlowGAN offers an efficient solution for unsupervised image-to-image translation by leveraging invertible flows.
• The method effectively reduces memory footprint and parameter redundancy, outperforming existing GANs in efficiency.
• iFlowGAN provides a generalizable framework for improving GAN-based image translation tasks.