Beyond Ground-Truth: Leveraging Image Quality Priors for Real-World Image Restoration

Fengyang Xiao; Peng Hu; Lei Xu; XingE Guo; Guanyi Qin; Yuqi Shen; Chengyu Fang; Rihan Zhang

cs.CVMar 31, 2026

Beyond Ground-Truth: Leveraging Image Quality Priors for Real-World Image Restoration

Authors: Fengyang Xiao, Peng Hu, Lei Xu, XingE Guo, Guanyi Qin, Yuqi Shen, Chengyu Fang, Rihan Zhang,

Organizations: Duke University · Tsinghua University · EPFL

Abstract

Real-world image restoration aims to restore high-quality (HQ) images from degraded low-quality (LQ) inputs captured under uncontrolled conditions. Existing methods typically depend on ground-truth (GT) supervision, assuming that GT provides perfect reference quality. However, GT can still contain images with inconsistent perceptual fidelity, causing models to converge to the average quality level of the training data rather than achieving the highest perceptual quality attainable. To address these problems, we propose a novel framework, termed IQPIR, that introduces an Image Quality Prior (IQP)-extracted from pre-trained No-Reference Image Quality Assessment (NR-IQA) models-to guide the restoration process toward perceptually optimal outputs explicitly. Our approach synergistically integrates IQP with a learned codebook prior through three key mechanisms: (1) a quality-conditioned Transformer, where NR-IQA-derived scores serve as conditioning signals to steer the predicted representation toward maximal perceptual quality. This design provides a plug-and-play enhancement compatible with existing restoration architectures without structural modification; and (2) a dual-branch codebook structure, which disentangles common and HQ-specific features, ensuring a comprehensive representation of both generic structural information and quality-sensitive attributes; and (3) a discrete representation-based quality optimization strategy, which mitigates over-optimization effects commonly observed in continuous latent spaces. Extensive experiments on real-world image restoration demonstrate that our method not only surpasses cutting-edge methods but also serves as a generalizable quality-guided enhancement strategy for existing methods. The code is available.

Figures & tables

Figure 2: Structural comparison of different training paradigms, where both (b) and our (c) utilize IQA to guide network training.

Figure 3: Results on blind face restoration. DifFace+ is the DifFace [ 59 ] with our quality prior conditioned approach, having more details. Our IQPIR has the highest perceptual quality.

Figure 4: Overall framework of IQPIR. (a) In the codebook learning stage, a dual-codebook architecture is proposed. The HQ+ codebook is learned to quantize $Z_{h}$ only when the quality score of $x_{h}$ is higher than the threshold $S_{thr}$ . (b) In the codebook lookup stage, we input the quality score $S$ as a condition into Transformer T , which predicts two code sequences at the same time. The two codebooks are leveraged to look up the corresponding code entries. Finally, the NR-IQA model is utilized to calculate the quality loss $\mathcal{L}_{quality}$ .

Data.	Methods	TOPIQ-G $\uparrow$	Musiq-G $\uparrow$	Musiq-K $\uparrow$	Musiq-A $\uparrow$	Arniqa $\uparrow$	MDFS $\downarrow$	Q-Align $\uparrow$	CLIP-IQA $\uparrow$
LFW-Test	RestoreFormer [ 50 ]	0.793	0.807	73.70	4.65	0.711	18.64	4.22	0.741
	DR2 [ 49 ]	0.720	0.753	67.13	4.67	0.700	18.43	4.23	0.658
	CodeFormer [ 67 ]	0.809	0.832	75.47	4.76	0.726	18.46	4.31	0.697
	DifFace [ 59 ]	0.718	0.748	69.90	4.48	0.686	18.41	3.86	0.610
	DAEFR [ 46 ]	0.814	0.827	75.84	4.81	0.742	17.91	4.33	0.696
	WaveFace [ 39 ]	0.786	0.799	73.55	4.69	0.689	19.18	4.43	0.788

Table 1: Results on blind face restoration. The best two results are in red and blue fonts, respectively.

		LOL-v1				LOL-v2-real				LOL-v2-synthetic
Methods	Sources	PSNR $\uparrow$	SSIM $\uparrow$	FID $\downarrow$	BIQE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	FID $\downarrow$	BIQE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	FID $\downarrow$	BIQE $\downarrow$
CUE [ 65 ]	ICCV23	21.86	0.841	69.83	27.15	21.19	0.829	67.05	28.83	24.41	0.917	31.33	33.83
GSAD [ 25 ]	NIPS23	23.23	0.852	51.64	19.96	20.19	0.847	46.77	28.85	24.22	0.927	19.24	25.76
AST [ 68 ]	CVPR24	21.09	0.858	87.67	21.23	21.68	0.856	91.81	25.17	22.25	0.927	37.19	28.78
MambaIR [ 12 ]	ECCV24	22.23	0.863	63.39	20.17	21.15	0.857	56.09	24.46	25.75	0.937	19.75	20.37
Reti-Diff [ 13 ]	ICLR25	25.35	0.866	49.14	17.75	22.97	0.858	43.18	23.66	27.53	0.951	13.26	15.77

Table 2: Results on low-light image enhancement.

		UIEB
Methods	Sources	PSNR $\uparrow$	SSIM $\uparrow$	UCIQE $\uparrow$	UIQM $\uparrow$
PUGAN [ 4 ]	TIP23	23.05	0.897	0.608	2.902
ADP [ 66 ]	IJCV23	22.90	0.892	0.621	3.005
NU2Net [ 11 ]	AAAI23	22.38	0.903	0.587	2.936
AST [ 68 ]	CVPR24	22.19	0.908	0.602	2.981
MambaIR [ 12 ]	ECCV24	22.60	0.916	0.617	2.991

Table 3: Results on underwater image enhancement.

Figure 5: Visualizations on low light, underwater and backlit image restoration.

		BAID
Methods	Sources	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
CLIP-LIT [ 33 ]	ICCV23	21.13	0.853	0.159	37.30
DiffIR [ 53 ]	ICCV23	21.10	0.835	0.175	40.35
AST [ 68 ]	CVPR24	22.61	0.851	0.156	32.47
MambaIR [ 12 ]	ECCV24	23.07	0.874	0.153	29.13
RAVE [ 8 ]	ECCV24	21.26	0.872	0.096	64.89

Table 4: Results on backlit image enhancement.

Table 9

Table 11

Geometry-Guided Image Restoration

See all 16 papers

QualiTeacher: Quality-Conditioned Pseudo-Labeling for Real-World Image Restoration
Mar 9, 2026Fengyang Xiao, Jingjia Feng, Peng Hu +6Image Restoration
Duke University · Tsinghua University · École Polytechnique Fédérale de Lausanne (EPFL) +2
Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration
Mar 27, 2026I-Hsiang Chen, Isma Hadji, Enrique Sanchez +5Image Restoration All-in-One Image Restoration
Samsung AI Center Cambridge · National Taiwan University · Technical University of Iasi +3
GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration
May 29, 2026Xiangtao Kong, Jixin Zhao, Lingchen Sun +2Synthetic Data Augmentation Image Restoration
The Hong Kong Polytechnic University · OPPO Research Institute

Mar 9, 2026J/K move · Enter open · S save

QualiTeacher: Quality-Conditioned Pseudo-Labeling for Real-World Image Restoration

Fengyang Xiao, Jingjia Feng, Peng Hu, Dingming Zhang+5

Duke University · Tsinghua University · École Polytechnique Fédérale de Lausanne (EPFL)+2

Image Restoration

Open paper AskPDF arXiv

Real-world image restoration (RWIR) is a highly challenging task due to the absence of clean ground-truth images. Many recent methods resort to pseudo-label (PL) supervision, often within a Mean-Teacher (MT) framework. However, these methods face a critical paradox: unconditionally trusting the often imperfect, low-quality PLs forces the student model to learn undesirable artifacts, while discarding them severely limits data diversity and impairs model generalization. In this paper, we propose QualiTeacher, a novel framework that transforms pseudo-label quality from a noisy liability into a conditional supervisory signal. Instead of filtering, QualiTeacher explicitly conditions the student model on the quality of the PLs, estimated by an ensemble of complementary non-reference image quality assessment (NR-IQA) models spanning low-level distortion and semantic-level assessment. This strategy teaches the student network to learn a quality-graded restoration manifold, enabling it to understand what constitutes different quality levels. Consequently, it can not only avoid mimicking artifacts from low-quality labels but also extrapolate to generate results of higher quality than the teacher itself. To ensure the robustness and accuracy of this quality-driven learning, we further enhance the process with a multi-augmentation scheme to diversify the PL quality spectrum, a score-based preference optimization strategy inspired by Direct Preference Optimization (DPO) to enforce a monotonically ordered quality separation, and a cropped consistency loss to prevent adversarial over-optimization (reward hacking) of the IQA models. Experiments on standard RWIR benchmarks demonstrate that QualiTeacher can serve as a plug-and-play strategy to improve the quality of the existing pseudo-labeling framework, establishing a new paradigm for learning from imperfect supervision. Code will be released.

Abstract

Geometry-Guided Image Restoration

Explore similar work

QualiTeacher: Quality-Conditioned Pseudo-Labeling for Real-World Image Restoration

Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration

GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration