Spatial-frequency hybrid Mamba with structure-aware scanning for face super-resolution

Face super-resolution (FSR) aims to restore clear and realistic high-resolution face images from low-resolution ones, which demands simultaneous preservation of facial symmetry, global proportion and fine-grained details. CNN-based methods suffer from limited receptive fields which lead to geometric distortion, while Transformers rely on global self-attention that tends to smooth identity-critical high-frequency details. The recently proposed Mamba employs selective state-space updates to model long-range dependencies, yet its native 1-D scanning breaks 2-D facial adjacency and lacks frequency-aware modeling, easily yielding asymmetric features and missing high-frequency details. To address these problems, we propose HybriMamba, a face super-resolution framework that equips Mamba with an Image-Selective Scan Module (Image-SSM) that restores pixel adjacency and enforces horizontal facial symmetry to maintain facial structure consistency, a Fourier–Wavelet Transform Module (FWM) that globally modulates magnitude spectra and directionally processes wavelet sub-bands to recover fine facial details, and a Local Enhancement Module (LEM) with pixel-wise gating that adaptively re-weights smooth and textured regions to avoid over-smoothing. Embedded in a U-shaped multi-scale encoder–decoder, HybriMamba hierarchically retains facial geometry while injecting micro-textures through skip connections. Extensive experiments on Helen, CelebA and the real-world SCFace dataset show that HybriMamba achieves optimal performance among state-of-the-art methods in terms of pixel-level fidelity, perceptual quality and identity consistency, delivering robust, identity-faithful and visually pleasing results under both synthetic and realistic degradations. Our code is released at https://github.com/LateAutu/HybriMamba.