One-Step Face Restoration via Shortcut-Enhanced Coupling Flow

TL;DR

This work establishes a data-dependent coupling that explicitly models the LQ--HQ dependency, minimizing path crossovers and promoting near-linear transport and achieves state-of-the-art one-step face restoration quality with inference speed comparable to traditional non-diffusion methods.

Abstract

Face restoration has advanced significantly with generative models like diffusion models and flow matching (FM), which learn continuous-time mappings between distributions. However, existing FM-based approaches often start from Gaussian noise, ignoring the inherent dependency between low-quality (LQ) and high-quality (HQ) data, resulting in path crossovers, curved trajectories, and multi-step sampling requirements. To address these issues, we propose Shortcut-enhanced Coupling flow for Face Restoration (SCFlowFR). First, it establishes a \textit{data-dependent coupling} that explicitly models the LQ--HQ dependency, minimizing path crossovers and promoting near-linear transport. Second, we employ conditional mean estimation to obtain a coarse prediction that refines the source anchor to tighten coupling and conditions the velocity field to stabilize large-step updates. Third, a shortcut constraint supervises average velocities over arbitrary time intervals, enabling accurate one-step inference. Experiments demonstrate that SCFlowFR achieves state-of-the-art one-step face restoration quality with inference speed comparable to traditional non-diffusion methods.

Figures (4)

• Figure 1: (a) Linear interpolation under independent marginal assumptions yields intersecting paths. (b)Vanilla FM learns a highly curved velocity field due to lack of source–target dependency modeling. (c) Coupling FM explicitly models dependencies between source and target, producing straighter trajectories. (d) Our SCFlowFR leverages shortcut constraints to learn the averaged velocity between arbitrary time steps, enabling one-step sampling.
• Figure 2: Overview of the SCFlowFR. We first construct a coupled transport path between the LQ and HQ image distributions (left), and leverage the preliminarily restored image from $\tau_\phi$ as conditional information $\bm{c}$ to guide the learning of velocity field $\bm{v_{\theta}}$. Furthermore, a shortcut constraint is introduced, enabling the model to predict the average velocity over a time interval $\Delta t$.
• Figure 3: Qualitative comparisons on the CelebA-Test dataset.
• Figure 4: Qualitative comparisons on three wild datasets. From top to bottom: samples from the CelebChild, LFW, and WebPhoto datasets, respectively.