RestoRect: Degraded Image Restoration via Latent Rectified Flow & Feature Distillation

TL;DR

RestoRect tackles degraded image restoration by bridging accuracy and efficiency through a generative knowledge-distillation framework based on latent rectified flow. The teacher–student pair leverages Retinex priors, Spatial Channel Layer Normalization, and a FLEX loss to align multi-scale transformer representations across heterogeneous architectures, enabling robust feature synthesis with few steps. The student learns velocity predictors for reflectance and image features under a two-phase training protocol, achieving diffusion-like restoration with as few as 4 inference steps and demonstrating state-of-the-art performance across 15 datasets and 4 tasks. This approach offers a scalable pathway for fast, high-fidelity image restoration and establishes a generalizable paradigm for cross-architecture knowledge transfer in vision models.

Abstract

Current approaches for restoration of degraded images face a trade-off: high-performance models are slow for practical use, while fast models produce poor results. Knowledge distillation transfers teacher knowledge to students, but existing static feature matching methods cannot capture how modern transformer architectures dynamically generate features. We propose a novel Latent Rectified Flow Feature Distillation method for restoring degraded images called \textbf{'RestoRect'}. We apply rectified flow to reformulate feature distillation as a generative process where students learn to synthesize teacher-quality features through learnable trajectories in latent space. Our framework combines Retinex decomposition with learnable anisotropic diffusion constraints, and trigonometric color space polarization. We introduce a Feature Layer Extraction loss for robust knowledge transfer between different network architectures through cross-normalized transformer feature alignment with percentile-based outlier detection. RestoRect achieves better training stability, and faster convergence and inference while preserving restoration quality, demonstrating superior results across 15 image restoration datasets, covering 4 tasks, on 10 metrics against baselines.

Figures (9)

• Figure 1: RestoRect achieves superior performance on four image restoration tasks while keeping parameter count low
• Figure 2: Training framework flowchart for RestoRect. Starting from top left (1. Prior Encoding) the inputs go through retinex decomposition and pass through encoders (1.1 ResNet Encoders) to prepare image and retinex prior encodings. Next these prior encodings are pre-trained (2. Teacher Pre-training) with the teacher model (2.2 UNet Transformer with SCLN) using a reconstruction loss. Finally the frozen prior encodings and teacher model are used for student phase 1 and phase 2 training using rectified flow loss. Full architecture details in Appendix \ref{['sec:arch']}.
• Figure 3: (1-4) Teacher model training with ablations of \ref{['ac']} & \ref{['ac']} Norm (red) and auxiliary losses (blue). Model indicated by red line without \ref{['ac']} & \ref{['ac']} Norm is identical to Reti-Diff he2023reti. (5) FID vs Steps inference performance show Rectified Flow (RF) student model producing high quality images in fewer steps compared to Denoising Diffusion Implicit Model (\ref{['ac']}).
• Figure 4: \ref{['ac']} task visual results (Top to Bottom: \ref{['ac']}-v1, v2-real, v2-syn, \ref{['ac']})
• Figure 5: \ref{['ac']} (Top), \ref{['ac']} (Middle Left), \ref{['ac']} (Middle Right), \ref{['ac']} (Bottom) task visual results