Search papers, labs, and topics across Lattice.
This paper introduces GeoMAR, a novel framework for blind face restoration that enhances feature conditioning through geometrically aligned features and a masked autoregressive (MAR) refinement process. By employing a dual-input extraction pipeline and an Aligned Geometric Priors Injector, GeoMAR effectively integrates spatially faithful anchors with low-quality features to improve robustness against severe degradation. Experimental results on multiple benchmarks show that GeoMAR significantly outperforms existing methods in terms of perceptual quality and visual coherence.
GeoMAR achieves superior face restoration by leveraging geometrically aligned features, outperforming traditional methods even under severe degradation.
Codebook-based blind face restoration (BFR) often suffers from ambiguous conditioning features and a fragile prediction mechanism under severe degradation. To address these challenges, we propose GeoMAR, a framework designed to unleash geometrically aligned features with masked autoregressive (MAR) refinement for robust face restoration. For feature conditioning, we introduce a dual-input extraction pipeline to extract component-based geometric descriptions with explicit, spatially faithful anchors. These textual priors are integrated with low-quality (LQ) features via an Aligned Geometric Priors Injector, which employs a KV-Q exchange strategy to generate geometrically aligned features. For prediction mechanism, we reformulate the one-step mapping into a multi-step MAR process. This coarse-to-fine generation progressively refines complex facial regions based on increasingly reliable context. Experiments on one synthetic and three real-world benchmarks demonstrate that GeoMAR achieves highly competitive perceptual quality and coherent visual structures compared with existing methods. The code is available at https://github.com/BRL-SYSU/GeoMAR.git.
