HQFace: Towards Real-World Blind Face Restoration with Hierarchical Vector Quantization

The goal of blind face restoration (BFR) is to reconstruct detailed and lifelike facial images from inputs that have undergone unknown forms of degradation. While current methods utilize priors to refine details and have achieved notable advancements, they often fail to fully utilize the intricate information within facial features and their interactions with these priors, leading to suboptimal results. Moreover, the discrepancy between synthetic simulations and actual-world conditions is frequently overlooked, compromising robustness and real-world efficacy. Our study introduces HQFace, a novel approach employing Hierarchical Vector Quantization (HVQ) to overcome these shortcomings, enhancing the capture and manipulation of contextual data with priors and integrating gated cross attention (GCA) for higher-quality, more realistic images. HQFace advances beyond traditional methods by using HVQ across multi-scale features for a detailed examination of spatial dynamics between degraded data and high-quality priors, ensuring restorations of greater realism and fidelity. Furthermore, our development of GCA targets superior facial components alignment with restoration goals, and an expanded degradation model encompasses a broader range of realistic scenarios, enhancing robustness and applicability. Extensive testing confirms HQFace’s superior performance over leading algorithms on the CelebA-HQ.