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

Researchers have proposed MaCo-GAN, a novel framework for Single Image Super-Resolution (SISR) that replaces conventional adversarial loss with a supervised contrastive objective, improving the perception-distortion trade-off.

Conventional Generative Adversarial Networks (GANs) for SISR often struggle with hallucinated artifacts because standard discriminators evaluate overall image naturalness rather than strict conditional realism, according to a paper submitted to arXiv on June 3, 2026^[1]. MaCo-GAN addresses this by introducing a dynamic fake sample synthesizer that transforms ground truth data into challenging, perceptually plausible fake images. The generator is trained to attract its predictions toward on-manifold fakes and repel them from off-manifold fakes, while the discriminator optimizes the opposite, as described in the paper. Another study on arXiv, submitted on March 23, 2026, and revised on June 5, 2026, highlights the challenges of hyperspectral image super-resolution, where existing methods focus solely on multispectral images and supervised deep learning approaches rely on accurate training data, which is often unavailable^[2]. The MaCo-GAN framework demonstrates consistent improvements across various benchmarks and has been validated through extensive ablation studies.