Bridging Information Asymmetry: A Hierarchical Framework for Blind Face Restoration with Reduced Uncertainty
Researchers have introduced Pref-Restore, a novel hierarchical framework designed to improve blind face restoration by reducing uncertainty. This framework utilizes an auto-regressive semantic branch to process image and text cues into structured tokens, providing a high-level anchor for the restoration process. A diffusion generator then uses this anchor to recover identity-relevant details, with a face-aware reward mechanism further refining the output to control the trade-off between quality and fidelity. Experiments indicate that Pref-Restore achieves state-of-the-art performance, demonstrating stronger identity-sensitive fidelity and lower restoration uncertainty compared to existing methods. AI