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

2025-09-27

AI Summary

Image restoration has long suffered from a fundamental trade-off between performance and efficiency: state-of-the-art models achieve high fidelity at the cost of slow inference, while lightweight models sacrifice quality for speed. To address this, we propose a novel framework grounded in latent-space correction flows and dynamic feature distillation, where knowledge transfer is formulated as generative latent trajectory learning. Our approach integrates Retinex decomposition, learnable anisotropic diffusion, and triangular color-space polarization to significantly enhance dynamic representational capacity in distillation. We further introduce cross-normalization Transformer-based feature alignment, hierarchical feature extraction losses, and physics-guided decomposition, augmented by percentile-based anomaly detection for improved robustness. Extensive evaluation across 15 datasets, 4 image restoration tasks, and 8 metrics demonstrates consistent superiority over SOTA methods. Moreover, our framework exhibits enhanced training stability, faster convergence, and accelerated inference.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Feature Construction/Reformulation
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Current approaches for restoration of degraded images face a critical trade-off: high-performance models are too 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 'RestoRect', a novel Latent Rectified Flow Feature Distillation method for restoring degraded images. 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 theory for physics-based 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. We demonstrate superior results across 15 image restoration datasets, covering 4 tasks, on 8 metrics.

Problem

• Addresses trade-off between restoration quality and inference speed
• Proposes latent rectified flow for dynamic feature distillation
• Enables cross-architecture knowledge transfer for image restoration

Innovation

• Latent rectified flow for feature distillation
• Retinex theory with anisotropic diffusion constraints
• Cross-normalized transformer feature alignment loss