PASDiff: Physics-Aware Semantic Guidance for Joint Real-world Low-Light Face Enhancement and Restoration

Abstract

Face images captured in real-world low light suffer multiple degradations-low illumination, blur, noise, and low visibility, etc. Existing cascaded solutions often suffer from severe error accumulation, while generic joint models lack explicit facial priors and struggle to resolve clear face structures. In this paper, we propose PASDiff, a Physics-Aware Semantic Diffusion with a training-free manner. To achieve a plausible illumination and color distribution, we leverage inverse intensity weighting and Retinex theory to introduce photometric constraints, thereby reliably recovering visibility and natural chromaticity. To faithfully reconstruct facial details, our Style-Agnostic Structural Injection (SASI) extracts structures from an off-the-shelf facial prior while filtering out its intrinsic photometric biases, seamlessly harmonizing identity features with physical constraints. Furthermore, we construct WildDark-Face, a real-world benchmark of 700 low-light facial images with complex degradations. Extensive experiments demonstrate that PASDiff significantly outperforms existing methods, achieving a superior balance among natural illumination, color recovery, and identity consistency.

Figures (24)

• Figure 1: Visual comparison on degraded synthetic and real-world inputs (a). Cascaded paradigms (b) L-Diff jiang2024lightendiffusion$\to$DiffBIR lin2024diffbir and (c) DiffBIR lin2024diffbir$\to$L-Diff jiang2024lightendiffusion suffer from severe noise amplification and structural collapse, respectively. Generic joint models like DarkIR feijoo2025darkir (d) and FDN tu2025fourier (e) struggle with complex degradations, leading to residual blur and detail loss. (f) Our PASDiff achieves superior perceptual quality with crisp facial details and natural illumination, effectively suppressing artifacts.
• Figure 1: Face recognition accuracy comparisons.
• Figure 2: From the phase reconstructions and statistical metrics, it can be seen that end-to-end learning strategies lin2024diffbir inherently adhere to the degraded intensity distribution, suffering from compromised facial identities and severe color shifts. Existing guidance approaches lin2025aglldiff correct global illumination and chromaticity, but lack semantic guidance for fine geometries. In contrast, PASDiff elegantly integrates physical and structural guidance to restore crisp textures, natural illumination, and a color distribution better aligned with the high-quality reference, effectively preserving facial identity.
• Figure 3: Overall framework of the proposed training-free PASDiff. Our method reformulates joint restoration via dual-dimensional diffusion guidance. Physically, Retinex-based photometric constraints steer the sampling trajectory to recover natural illumination and chromaticity. Structurally, our Style-Agnostic Structural Injection (SASI) and Statistic-Aligned Guidance Loss extract high-frequency facial semantics from priors while explicitly filtering out their intrinsic lighting biases. This synergy seamlessly harmonizes texture realism with physical reliability.
• Figure 4: Left: Visualization of physical priors (Exposure and Reflectance maps). Right: Subjective preferences from the user study.