In this paper, we present a novel nonlocal image denoising model that integrates a threshold-based diffusion function with a fidelity term defined in the negative Sobolev space \( H^{-s} \) , where \( s \in (0,1] \) . The use of the \( H^{-s} \) norm proves particularly effective in preserving fine textures and subtle structures that are often lost in traditional denoising approaches. The diffusion mechanism is inspired by Perona–Malik-type models and is adapted using a smoothed gradient of the input image, enhancing edge preservation while reducing noise. Experimental results demonstrate that our method consistently outperforms several state-of-the-art denoising techniques in terms of both visual quality and quantitative metrics, confirming its robustness and effectiveness.