RestoRect: Degraded Image Restoration via Latent Rectified Flow & Feature Distillation
RestoRect proposes a new method called Latent Rectified Flow Feature Distillation to restore degraded images. It combines Retinex decomposition with learnable diffusion constraints and trigonometric color space polarization. The student network learns to synthesize high-quality features by following learned trajectories in latent space, leading to faster convergence and better restoration quality across various image tasks.