Image Restoration with Conditional Noise Diffusion: Restoring Facial Clarity and Details

Restoring facial structures in degraded images is a challenging task in image processing due to issues like noise, blur, and loss of fine details. This paper introduces Diffusion-GFPGAN, a novel approach that integrates a custom-built diffusion model with GFPGAN to enhance facial restoration. The custom model is designed for denoising and noise handling, while GFPGAN excels in feature extraction and image enhancement. By leveraging both models’ strengths through multi-scale feature fusion, the proposed methodology significantly improves the restoration of facial details. Extensive experiments conducted on the CelebA-Test dataset, which includes diverse facial variations, demonstrate that Diffusion-GFPGAN outperforms state-of-the-art models in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). The model achieves PSNR scores of 25.91 dB and SSIM scores of 0.6952, indicating improved restoration quality and visual realism. This approach presents a powerful solution for restoring degraded facial images while offering a promising framework for broader image restoration applications.