Title: MaCo-GAN: Manifold-Contrastive Adversarial Learning for Single Image Super-Resolution

Abstract: Standard Generative Adversarial Networks (GANs) employed in Single Image Super-Resolution (SISR) frequently suffer from hallucinated artifacts. This issue stems from the tendency of traditional discriminators to assess general image naturalness rather than ensuring strict conditional realism. To overcome these limitations, we introduce MaCo-GAN, an innovative manifold-contrastive GAN architecture that substitutes the standard adversarial loss with a supervised contrastive objective. A pivotal element of our approach is a dynamic fake sample synthesizer, which converts ground truth (GT) data into a diverse range of challenging yet perceptually realistic fake images that preserve low-resolution (LR) correspondence. Leveraging these synthesized samples, we formulate a robust contrastive minimax game. In this setup, the generator is trained to pull its predictions toward on-manifold fakes (characterized by low distortion) while pushing them away from off-manifold fakes (high distortion); the discriminator, conversely, is optimized to perform the opposite actions. By merely swapping the adversarial loss in a baseline SR model with our proposed objective, we achieve consistent enhancements in the perception-distortion trade-off across multiple benchmarks. Comprehensive ablation studies confirm the efficacy of our framework and offer profound insights into the dynamics of this conditional contrastive game.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC