TDiff: Thermal Plug-And-Play Prior with Patch-Based Diffusion

Thermal images captured by low-cost sensors often suffer from low resolution, fixed pattern noise, and other localized degradations. Compounding this challenge, available thermal imaging datasets are typically small and lack diversity.

To address these limitations, we introduce TDiff, a plug-and-play thermal restoration framework built on a patch-based diffusion prior. Our approach exploits the inherently local structure of thermal distortions by training diffusion models on small patches, then restoring full-resolution images by denoising overlapping patches and blending them using smooth spatial windowing.

To our knowledge, this is the first diffusion-based patch prior designed specifically for thermal imagery and applicable across multiple restoration tasks. Experiments on denoising, super-resolution, and deblurring demonstrate strong performance on both simulated and real thermal datasets, establishing TDiff as a unified and effective pipeline for thermal image restoration.