Zero-Shot Blind Face Restoration Via Conditional Diffusion Sampling

Haowei Li, Dongyu Zhang

Abstract: Blind Face Restoration (BFR) aims to restore high-quality faces from low-quality input faces without knowledge of the degradation model. The key challenge is maintaining image fidelity while introducing generation priors to enhance realness. In recent years, research has focused on leveraging the generative prior of diffusion models to restore degraded images. However, existing methods face two main limitations: (1) Non-zero-shot methods require substantial paired training data. During training, a predefined degradation model is usually used to synthesize degraded low-quality images, which makes the algorithm less robust to rare degradations. (2) Balancing realness and fidelity is challenging. We propose a two-stage BFR approach. First, we employ a diffusion model that combines a conditional control strategy based on discrete wavelet transform (DWT) and a steered diffusion algorithm for the conditional constraint to remove the degradation. Next, we utilize an image enhancement model to improve image quality. Both stages employ pre-trained models without additional training. This method achieves outstanding results in terms of fidelity-realism trade-offs, outperforming existing zero-shot restoration models.