GeoMAR: Unleashing Geometrically Aligned Features for Masked Autoregressive Blind Face Restoration

GeoMAR targets blind face restoration under severe degradation. Its dual-input pipeline extracts component-level geometric descriptions with explicit spatial anchors, then combines these textual priors with low-quality image features through an Aligned Geometric Priors Injector using a KV-Q exchange strategy. The method also replaces one-step code prediction with multi-step masked autoregressive refinement, progressively reconstructing difficult facial regions from increasingly reliable context. The authors report competitive perceptual quality and coherent structures on one synthetic and three real-world benchmarks. Code is linked on GitHub, although the supplied abstract does not include numerical results.