Towards Universal Image Translation: A Dual-Mode GAN Approach for Visual Modalities

In this paper, we propose a new dual-mode generative adversarial network (GAN) architecture that is capable of performing efficient and adaptive image translation between visual modalities, with a focus on grayscale-to-colour and thermal-to-visible conversions. In contrast to conventional models that tend to be specialized for one task or modality, our method uses a single, lightweight encoder that maps both grayscale and thermal images to a common latent space. This facilitates smooth translation across modalities without increasing computational overhead. The architecture combines expert generators and discriminators tailored to the target modality, producing high-quality output with adversarial, reconstruction, and consistency losses. Our model is optimized for efficiency and can handle real-time requirements as well as deployment on constrained platforms. Large-scale experiments and ablation tests prove the efficiency of our approach, presenting competitive performance on conventional image quality measures like PSNR, SSIM, and FID. This paper is an advance toward general image translation systems that can work with a wide variety of input types under a unified, coherent framework.[1][3]