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

Shourya Verma; Mengbo Wang; Nadia Atallah Lanman; Ananth Grama

cs.CVSep 27, 2025

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

Authors: Shourya Verma, Mengbo Wang, Nadia Atallah Lanman, Ananth Grama

Organizations: Department of Computer Science, Purdue University, USA · Purdue Institute of Cancer Research, USA

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 '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 & tables

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 (2.1 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 B.7 .

Figure 3: (1-4) Teacher model training with ablations of SCLN & QK Norm (red) and auxiliary losses (blue). Model indicated by red line without SCLN & QK Norm is identical to Reti-Diff ( He et al., 2023 ) . (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 ( DDIM ).

Figure 4: Comparison of normalization strategies for image restoration transformers. Evaluated SCLN against LayerNorm (per-token, the baseline used in Reti-Diff), RMSNorm (per-token, no mean subtraction), and i-LN ( ( Lee et al., 2025 ) , norm with global-std rescaling)

Methods	LOL -v1				LOL -v2-real				LOL -v2-syn				SID
Methods	PSNR $\uparrow$	SSIM $\uparrow$	FID $\downarrow$	BIQI $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	FID $\downarrow$	BIQI $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	FID $\downarrow$	BIQI $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	FID $\downarrow$	BIQI $\downarrow$
MRQ ( Liu et al., 2023 )	25.24	0.855	53.32	22.73	22.37	0.854	68.89	33.61	25.54	0.940	21.56	25.09	24.80	0.688	63.72	29.53
IAGC ( Wang et al., 2023b )	24.53	0.866	59.73	25.50	22.20	0.863	70.34	31.70	25.58	0.941	21.58	30.32	23.17	0.640	78.80	30.56
DiffIR ( Xia et al., 2023 )	23.15	0.828	70.13	26.38	21.15	0.816	72.33	29.15	24.76	0.921	21.36	27.74	23.17	0.640	78.80	30.56
CUE ( Zheng et al., 2023 )	21.86	0.841	69.83	27.15	21.19	0.829	67.05	28.83	24.41	0.917	31.34	33.83	23.25	0.652	77.38	28.85
Mamba ( Guo et al., 2024 )	22.33	0.863	63.39	20.17	21.97	0.840	56.09	24.46	25.75	0.958	17.95	20.37	21.14	0.656	154.76	32.72

Table 1: LLIE task results. Best result shown in Green and second best shown in Blue.

Figure 5: LLIE task visual results (Top to Bottom: LOL -v1, v2-real, v2-syn, SID )

Table 7

Figure 6: FIE (Top), UIEB (Middle Left), LSUI (Middle Right), BAID (Bottom) task visual results

Table 9

Figure 7: DCIM (Row 1 Left), LIME (Row 1 Right), MEF (Row 2 Left), NPE (Row 2 Right), VV (Row 3), SICE -Grad (Row 4) results

Figure 8: LLIE task visual results with ablation of SCLN and QK Norm, Retinex Priors, FLEX loss, compared to full RestoRect architecture and Ground Truth.

Table 12

Figure 9: Failure cases on SICE -Mix dataset.

Configuration	PSNR $\uparrow$	SSIM $\uparrow$	FID $\downarrow$	BIQI $\downarrow$
RetiDiff (Traditional KD)	25.35	0.866	49.14	17.75
1-Vel-Net RF(TC) + no FLEX	21.10	0.762	52.19	20.95
1-Vel-Net RF(TC) + FLEX	21.52	0.791	51.24	20.88
RF(no TC) + no FLEX	24.10	0.855	49.28	17.95
RF(TC) + no FLEX	24.27	0.891	44.75	9.02
RF(TC) + FLEX(CN)	25.64	0.913	42.18	8.73

Table 10: Student ablation study (L-v1).

Figure 10: Qualitative human evaluation user study with 17 participants on LLIE datasets.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Acronym	Full Form
Tasks & Methods
LLIE	Low-Light Image Enhancement
UIE	Underwater Image Enhancement
FIE	Fundus Image Enhancement
SCLN	Spatial Channel Layer Normalization
FLEX	Feature Layer EXtraction Loss

Appendix

Table 11: List of acronyms used in this paper

Dataset	PSNR	SSIM	FID	NIQE	LPIPS	BRISQ	BIQI	UCIQE	UIQM	CLIPQ
LOL -v1	27.85	0.94	38.67	7.47	0.11	27.16	8.35	0.52	2.60	0.499
LOL -v2 Real	22.97	0.91	42.81	7.74	0.13	28.44	10.48	0.51	2.88	0.500
LOL -v2 Syn	27.70	0.97	16.75	5.74	0.06	15.68	11.68	0.55	2.77	0.498
SID	26.19	0.92	54.23	5.87	0.15	20.05	19.57	0.85	2.38	0.498
UIEB	25.89	0.95	20.26	6.49	0.11	17.89	13.99	0.58	3.12	0.501
LSUI	28.10	0.94	17.83	5.04	0.18	21.82	16.06	0.57	3.23	0.499

Appendix

Table 12: RestoRect Image Quality Evaluation Results for all 15 datasets across 10 metrics

Image Restoration

Image restoration uses computational methods to recover useful visual content from noisy, blurred, low-resolution, or otherwise degraded images.

See all 364 papers

Flow-Map Distillation on Relation Manifolds for Image Restoration
Aug 6, 2026Zihao He, Songhua LiuFlow Matching Flow Map Learning
School of Artificial Intelligence · Shanghai Jiao Tong University · Shanghai City, China
IR-Flow: Bridging Discriminative and Generative Image Restoration via Rectified Flow
Apr 21, 2026Zihao Fan, Xin Lu, Jie Xiao +3Rectified Flow All-in-One Image Restoration
MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, School of Information Science and Technology, University of Science and Technology of China
ScaleResfusion: Residual Rectified Flow based on Residual Vector Field
Jul 28, 2026Zhenning Shi, Chen Xu, Junhao Zhang +4Diffusion Models for Image Restoration Rectified Flow
College of Computer Science, Nankai University, Tianjin, China · Faculty of Information Science and Computing, University of Macau, Macau, China · CSIRO Data61, Australia +1

Aug 6, 2026J/K move · Enter open · S save

Flow-Map Distillation on Relation Manifolds for Image Restoration

Zihao He, Songhua Liu

School of Artificial Intelligence · Shanghai Jiao Tong University · Shanghai City, China

Flow Matching Flow Map Learning Image Restoration Relational Knowledge Distillation Knowledge Distillation

Open paper AskPDF arXiv

Knowledge distillation for image restoration typically aligns intermediate features or relation matrices between teacher and student networks as static targets, ignoring the dynamic structure of the knowledge transfer process. In this paper, we propose Flow-Map Distillation on Relation Manifolds (FoRM), which reformulates relation-based knowledge transfer as a continuous flow mapping problem on the relation manifold. Rather than regressing a constant velocity field between student and teacher relation states, FoRM learns a flow map operator

\mathcal{F}_θ(\mathbf{z}, t, s)

that directly predicts the relation state at any target time

s

given the current state at time

t

, enabling richer trajectory-level supervision. To ensure global self-consistency of the learned flow map, we introduce a safe semigroup consistency constraint that enforces compositional agreement using ground-truth bridge states, eliminating phantom-state error accumulation. An endpoint anchoring loss further prevents the operator from drifting away from the teacher target. Extensive experiments on five image restoration tasks, including super-resolution, deraining, denoising, deblurring, and low-light enhancement, demonstrate consistent gains over state-of-the-art distillation baselines across multiple backbone architectures, reducing training variance by approximately 50% compared to naive flow matching distillation while achieving superior restoration quality.

Abstract

Image Restoration

Explore similar work

Flow-Map Distillation on Relation Manifolds for Image Restoration

IR-Flow: Bridging Discriminative and Generative Image Restoration via Rectified Flow

ScaleResfusion: Residual Rectified Flow based on Residual Vector Field