Exploring Diffusion-generated Guidance for Thermal Image Super-resolution

Leo Thomas Ramos, Angel D. Sappa; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2026, pp. 105-114

This work presents a strategy for guided thermal super-resolution that replaces the conventional RGB guidance with a pseudo-thermal image. The approach aims to enhance spectral alignment between the guidance and target images, improving the transfer of structural and thermal information during reconstruction. The pseudo-thermal guidance is generated using a diffusion model trained for thermal image synthesis and integrated into several state-of-the-art guided super-resolution methods. Experiments conducted on the M3FD and CIDIS datasets at x8 and x16 scaling factors demonstrate consistent performance gains over RGB-guided counterparts, including PSNR gains of up to +1.64 dB, SSIM improvements of +0.0290, and LPIPS reductions of 0.11 points. Visual results further confirm that the proposed strategy yields sharper details, cleaner contours, and perceptually superior reconstructions.