PixelSR: Pixel-Space Orthogonal Flow Decoupling for Real-World Image Super-Resolution

Xiang Lv; Kai Zhao; Qiang Wang; Qiao Mo; Chao Zhou; Mingwen Shao

cs.CVOct 20, 2026

PixelSR: Pixel-Space Orthogonal Flow Decoupling for Real-World Image Super-Resolution

Authors: Xiang Lv, Kai Zhao, Qiang Wang, Qiao Mo, Chao Zhou, Mingwen Shao

Organizations: China University of Petroleum (East China) · Kuaishou Technology · Shenzhen University of Advanced Technology

Abstract

Latent diffusion models have recently achieved remarkable progress in Real-world Image Super-Resolution (Real-SR). However, existing methods consistently exhibit a persistent seesaw effect: enhancing perceptual realism compromises structural fidelity. This effect, we argue, is not intrinsic to diffusion models, but largely stems from the architectural constraints of prevailing two-stage frameworks. To address this challenge, we propose PixelSR, a novel single-stage Pixel Diffusion Transformer for Real-SR that operates directly in pixel space. Specifically, PixelSR consists of two core designs: Pixel-space Conditional Diffusion and Orthogonal Flow Decoupling. The former establishes an effective pixel-space restoration framework by integrating spatially aligned LR conditioning, visual semantic guidance, and lightweight texture refinement, thereby bypassing VAE compression and enabling end-to-end optimization directly in RGB space. Complementarily, the latter decomposes the asymmetric flow target into a low-rank semantic component and a noise-free orthogonal texture component. By routing these components to specialized branches, it achieves decoupled optimization with reduced mutual interference. Benefiting from these designs, our PixelSR restores perceptually realistic details while faithfully preserving structural content. Extensive experiments across multiple benchmarks demonstrate that PixelSR not only achieves competitive performance, but also effectively mitigates the fidelity--realism seesaw effect.

Figures & tables

Figure 2: Overview of the proposed PixelSR framework. PixelSR comprises Pixel-space Conditional Diffusion (PCD) and Orthogonal Flow Decoupling (OFD). Left: PCD performs end-to-end RGB-space restoration using spatially aligned LR conditioning, visual semantic guidance, and lightweight pixel-level texture refinement. Right: OFD factorizes the asymmetric velocity target into a projected low-rank component and a deterministic orthogonal residual, which empirically exhibit semantic- and texture-dominant characteristics, respectively. Their predictions are assigned to specialized DiT and PiT branches and recombined into the full velocity for flow matching, enabling task-aligned optimization with reduced interference.

Figure 3: Visualization of orthogonal target factorization. With the Procrustes-aligned projector, the projected component mainly retains global content and coarse structures, while its orthogonal complement emphasizes edges, text patterns, and high-frequency textures.

Dataset	Method	Steps	Inference Time (s)	Reference-based Metrics				No-Reference Metrics				Rank $\downarrow$
Dataset	Method	Steps	Inference Time (s)	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	DeQA $\uparrow$	Rank $\downarrow$
	StableSR	200	5.75	20.4946	0.5227	0.3193	60.0564	4.6047	62.7178	0.3254	3.0055	9
	DiffBIR	50	5.08	20.5132	0.4899	0.2651	58.4458	4.4425	73.2548	0.6209	4.1696	4
	DiT4SR	40	5.26	19.2087	0.4804	0.2723	59.0254	4.2242	72.6805	0.5067	3.8741	8
	VOSR	25	3.86	20.5993	0.5111	0.2386	46.4918	3.8768	72.9454	0.5194	4.0873	2
LSDIR	DiT-SR	15	1.84	21.1127	0.5490	0.2457	41.7907	4.2975	69.5674	0.4341	3.8280	6

Table 1: Quantitative comparison on LSDIR, RealSR, ScreenSR, and UGCSR. Red indicates the best performance, orange indicates the second-best performance, and yellow indicates the third-best performance.

\uparrow

: higher is better;

\downarrow

: lower is better.

Figure 4: Qualitative comparison on challenging real-world examples. PixelSR better preserves the text layout and feather structure while recovering fine-grained details consistent with the HR references, whereas several competing methods exhibit distorted characters, irregular textures, or over-smoothed results. Best viewed zoomed in.

Figure 5: Ablation of PixelSR components on RealSR. The bars and green curve report PSNR and MANIQA, respectively. Pixel-space $\bm{x}_{0}$ -prediction provides a substantially stronger baseline than pixel-space $\bm{v}$ -prediction, while the successive PCD and OFD designs further improve fidelity and perceptual quality.

Figure 6: Frequency analysis of OFD branch specialization. Restoration results (top) and corresponding power spectra (bottom) under increasing orthogonal scale $\alpha$ . Larger $\alpha$ progressively strengthens fine textures and mid-to-high-frequency responses, while $\alpha=1$ denotes the default OFD output. All spectra share the same color scale.

Variant	Reference-based Metrics			No-Reference Metrics
Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	DeQA $\uparrow$
$\bm{v}$ -prediction	15.0173	0.0995	1.2483	26.5085	0.2176	1.5582
$\bm{x}_{0}$ -prediction	19.5442	0.4822	0.2543	72.3934	0.4895	3.6563
$\bm{u}_{\mathrm{A}}$ -prediction	20.3088	0.5101	0.2251	72.4436	0.5495	3.9879
OFD (Ours)	20.6907	0.5270	0.2203	73.1895	0.5661	4.0936

Table 2: Ablation of pixel-space flow formulations on LSDIR. All variants use the same PCD network and parameter count. OFD improves over monolithic

\bm{u}_{\mathrm{A}}

-prediction through explicit factorization and branch assignment.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Architecture
Hyperparameter	PixelSR
Restoration resolution	$512^{2}\times 3$
Conditional input dimension	$512^{2}\times 6$
Patch-level depth $N$	$28$
Pixel-level depth $M$	$4$
Hidden size $D_{\mathrm{DiT}}$	$1024$

Appendix

Table 3: Implementation details of PixelSR.

Figure 7: Representative image matrix of UGCSR benchmark. The benchmark covers ten semantic categories: humans, animals, text-rich images, natural scenes, urban scenes, indoor scenes, plants, food, tools, and everyday objects. Two representative degraded images are shown for each category.

Figure 8: Objective semantic analysis of UGCSR. (a) Distribution of OpenCLIP zero-shot pseudo-labels across nine semantic categories. (b) UMAP visualization of the OpenCLIP image features, where each point represents one HR image and colors indicate the corresponding Top-1 predicted categories.

Metric	RealSR		ScreenSR		UGCSR
Metric	LR	HR	LR	HR	LR	HR
MUSIQ $\uparrow$	29.597 $\pm$ 6.335	57.465 $\pm$ 9.239	41.220 $\pm$ 12.321	72.150 $\pm$ 4.672	69.753 $\pm$ 9.747	75.175 $\pm$ 3.196
MANIQA $\uparrow$	0.278 $\pm$ 0.084	0.329 $\pm$ 0.093	0.212 $\pm$ 0.081	0.565 $\pm$ 0.061	0.464 $\pm$ 0.112	0.588 $\pm$ 0.056
DeQA $\uparrow$	1.721 $\pm$ 0.186	3.050 $\pm$ 0.418	2.341 $\pm$ 0.403	4.121 $\pm$ 0.303	3.559 $\pm$ 0.537	4.183 $\pm$ 0.215
NIQE $\downarrow$	18.215 $\pm$ 28.351	6.117 $\pm$ 1.444	6.968 $\pm$ 1.248	3.772 $\pm$ 1.952	4.797 $\pm$ 1.841	3.884 $\pm$ 1.428

Appendix

Table 4: No-reference quality comparison of LR and HR images across real-world paired SR benchmarks. Results are reported as mean

\pm

standard deviation. The best results among LR and HR images are highlighted in bold and underlined , respectively.

Variant	Reference-based Metrics			No-Reference Metrics
Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	DeQA $\uparrow$
Learnable $\bm{P}$	23.4996	0.7114	0.2222	71.8488	0.4687	3.7217
SVD-based $\bm{P}$	23.5053	0.7010	0.2224	71.9188	0.4645	3.7063
Procrustes $\bm{P}$ (SD2.1)	23.5640	0.7156	0.2082	72.8613	0.4824	3.8294
Procrustes $\bm{P}$ (SD3)	23.5583	0.7160	0.2132	72.8709	0.4816	3.8146
Procrustes $\bm{P}$ (FLUX.2)	23.5661	0.7168	0.2099	72.8436	0.4809	3.8318

Appendix

Table 5: Ablation of projector construction on ScreenSR. The first three rows compare projection strategies, while the additional rows examine different source VAEs for Procrustes alignment. All variants use the same training settings. Red and orange indicate the best and second-best performance, respectively.

Figure 9: t-SNE visualization of semantic and texture features. We sample 300 images from each of five ImageNet categories and visualize features from the two PixelSR branches. Semantic features form compact category-aware clusters, while texture features are more dispersed due to diverse local details, supporting the effectiveness of our OFD.

Encoder	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	DeQA $\uparrow$
CLIP	25.2911	0.7161	0.2924	68.4463	0.5056	3.2331
SigLIP	25.1778	0.7099	0.2967	68.7191	0.5168	3.2396
DINOv2	25.3360	0.7166	0.2922	69.5941	0.5381	3.6116

Appendix

Table 6: Ablation on different semantic encoders on RealSR. All variants use the same PCD framework and settings. Red indicates the best performance.

Figure 10: Qualitative comparison on challenging real-world examples. PixelSR better preserves the original text layout and character structures while recovering fine-grained image details, whereas several competing methods exhibit distorted characters, irregular textures, or over-smoothed results. Best viewed zoomed in.

Figure 11: Qualitative comparison on challenging real-world examples. PixelSR better preserves high-frequency textures and fine structural details, whereas competing methods often produce over-smoothed surfaces, irregular patterns, or artificial artifacts. Best viewed zoomed in.

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 247 papers

PixelIR: Fidelity-Perception Decoupling via Pixel-Space Image-Residual Flow Matching for Efficient One-Step Real-World Super-Resolution
Aug 31, 2026Bingtian Qiao, Yue Shi, Yong Guo +2Image Super-Resolution
Shanghai Jiao Tong University · Fuzhou University · Shanghai AI Laboratory
FreqOrtho-SR: Frequency-Guided Orthogonal Expert Learning for Real-World Image Super-Resolution
Jun 27, 2026Minh Son Hoang, Dinh Phu Tran, Quyen Nguyen Duc +2Diffusion Models for Image Restoration Mixture of Experts
School of Computing, KAIST, Republic of Korea
Degradation-Aware and Structure-Preserving Diffusion for Real-World Image Super-Resolution
Apr 13, 2026Yang Ji, Zonghao Chen, Zhihao Xue +1Degradation-Aware Image Restoration Image Super-Resolution
KUNBYTE Hangzhou, Zhejiang Province, China · GOLDMYE Xuhui District, Shanghai, China

Aug 31, 2026J/K move · Enter open · S save

PixelIR: Fidelity-Perception Decoupling via Pixel-Space Image-Residual Flow Matching for Efficient One-Step Real-World Super-Resolution

Bingtian Qiao, Yue Shi, Yong Guo, Wenjun Zhang+1

Shanghai Jiao Tong University · Fuzhou University · Shanghai AI Laboratory

Image Super-Resolution

Open paper AskPDF arXiv

Real-world image super-resolution (Real-ISR) aims to preserve structures supported by the degraded observation while reconstructing perceptually realistic details. However, existing Real-ISR methods largely optimize fidelity and perceptual quality within a shared network, causing the two objectives to interfere throughout training and making their balance difficult to control. Recent one-step methods reduce sampling steps, yet often inherit both this coupled optimization behavior and the expensive high-resolution backbone of their multi-step predecessors. We argue that efficient Real-ISR requires not only a shorter sampling trajectory, but also specialized modeling of faithful reconstruction and perceptual detail synthesis. Based on this insight, we propose PixelIR, a fidelity-perception decoupling framework built upon pixel-space image-residual flow matching. PixelIR first learns an image flow that maps the degraded observation to a faithful reconstruction. Then, a residual flow synthesizes the missing perceptual details from noise without repeatedly relearning or overwriting the complete restoration solution. We further distill the teacher into a deployment-oriented one-step student within a coarse-to-fine pyramid architecture. Extensive experiments show that PixelIR achieves leading PSNR, SSIM, and LPIPS on both RealSR and DRealSR. The final model completes pixel-space restoration in a single evaluation with only 32.9M parameters, 89.7G MACs, and 8.5ms latency, demonstrating a strong practical fidelity-perception-efficiency balance.

Abstract

Image Super-Resolution

Explore similar work

PixelIR: Fidelity-Perception Decoupling via Pixel-Space Image-Residual Flow Matching for Efficient One-Step Real-World Super-Resolution

FreqOrtho-SR: Frequency-Guided Orthogonal Expert Learning for Real-World Image Super-Resolution

Degradation-Aware and Structure-Preserving Diffusion for Real-World Image Super-Resolution