cs.CVOct 10, 2026

DuSRFlow: Dual-Lens Super-Resolution via Parallax-Aware Flow Estimation and Complementary Reference Warping

Authors: Fangpu Zhang, Huanjing Yue, Zifan Cui, Shaoxiong Tu, Qirui Yang, Jingyu Yang

Organizations: State Key Laboratory of Smart Power Distribution Equipment and System, Tianjin University and School of Electrical and Information Engineering, Tianjin University · School of Electrical and Information Engineering, Tianjin University, Tianjin, China, 300072 · Huawei Technologies Co., Ltd.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Dual-lens super-resolution (DuSR) aims to enhance a wide-angle low-resolution (LR) image by exploiting a telephoto reference (Ref) image. A central challenge in DuSR is reliable reference utilization under large parallax, occlusion, and cross-camera imaging discrepancies. Existing methods often employ off-the-shelf monocular optical flow models to provide geometric priors. However, their assumptions of photometric consistency and smooth motion are violated in dual-lens scenarios, leading to unreliable alignment and degraded reference-based reconstruction. To address these issues, we propose DuSRFlow, a dual-lens super-resolution framework that improves reference utilization from both correspondence learning and reference aggregation. First, we introduce DuFlowNet, a dedicated dual-lens flow estimator for robust cross-view alignment. To train DuFlowNet without ground-truth flow annotations, we develop a real--synthetic data fusion strategy that composites synthetic foreground objects onto real dual-lens image pairs, generating controllable large-parallax samples while preserving realistic cross-camera imaging characteristics. Second, we observe that reference reliability is highly spatially uneven in DuSR, where unreliable local reference responses may suppress the utilization of reliable cues. We therefore propose a Complementary Reference Warping strategy that coordinates patch-based warping and kernel-free-guided deformable warping according to local reference reliability, producing a more reliable reference representation before LR--Ref fusion. In addition, we construct DuSR-RealV2, a challenging real-world benchmark with complex depth variations, large parallax, and occlusion. Extensive experiments on multiple DuSR benchmarks demonstrate that DuSRFlow achieves state-of-the-art performance in reconstruction quality, perceptual quality, and cross-dataset generalization.

Figures & tables

Fig. 1: (a) LR–Ref patches suffer from cross-camera discrepancies and parallax, challenging generic optical flow estimation. (b) Benefiting from real–synthetic data fusion, our dual-lens flow achieves cleaner alignment than SPyNet with fewer warping artifacts. (c) More accurate alignment improves reference utilization, leading to more faithful structures and fewer reference-imprinted artifacts than KeDuSR.
Fig. 2: Although the central reference regions have the same quality in both cases, the final SR results differ because unreliable corner references reduce the overall reliability of the fused reference representation.
Fig. 3: Distributions of Occlusion Rate (OR) and Average Flow Magnitude (AFM) across different DuSR datasets. DuSR-RealV2-Paired exhibits the lowest and most compact OR distribution, indicating highly reliable paired alignment. In contrast, DuSR-RealV2-Unpaired shows a broader distribution shifted toward larger AFM values, demonstrating stronger LR–Ref parallax and higher evaluation difficulty.
Fig. 4: The framework of our DuSRFlow. The proposed dual-lens flow based alignment module first aligns the reference image IRef\mathbf{I}^{\mathrm{Ref}} to the LR center region, yielding the aligned reference feature F^Ref\hat{\mathbf{F}}^{\mathrm{Ref}} . In parallel, kernel-free (KF) matching between ILR\mathbf{I}^{\mathrm{LR}} and ILRC\mathbf{I}^{\mathrm{LRC}} produces the KF index and confidence map, while the SISR encoder extracts the LR feature FLR\mathbf{F}^{\mathrm{LR}} . These features are then processed by the proposed complementary reference warping to generate the aggregated reference feature F~Ref\tilde{\mathbf{F}}^{\mathrm{Ref}} , which is fused with the LR feature through LR–Ref fusion to produce the final SR result ISR\mathbf{I}^{\mathrm{SR}} .
Fig. 5: (a) Illustration of dual-lens parallax and view-dependent occlusion. A foreground object can exhibit different projections and occlude different background regions across views. (b) The proposed real–synthetic data fusion strategy. A synthetic foreground is composited onto real dual-lens image pairs, where different homography transformations are applied to the reference branch and the LR/HR-center branches to generate controllable parallax while preserving realistic dual-lens imaging characteristics.
Fig. 6: A toy experiment to evaluate DuFlowNet. (a) HR image reconstruction with pixel-domain patch matching using different matching methods. (b) Patch similarity and hit rate curves with different matching methods.
Fig. 7: Validation of KF confidence as a proxy for cross-camera reference-transfer reliability. All valid patches are divided into 20 equal-count bins according to KF confidence CC , and each point reports the mean CC and mean ideal reference confidence CidealC^{\mathrm{ideal}} within each bin. The Spearman correlation over all patches is 0.6370.637 .
MethodFull-ImageCenter-ImageCorner-Image
PSNR ↑\uparrow /SSIM ↑\uparrow /LPIPS ↓\downarrowPSNR/SSIMPSNR/SSIM
RCAN- ℓ\ell27.16 / 0.8256 / 0.26627.77 / 0.829827.02 / 0.8242
RCAN26.25 / 0.7866 / 0.14526.85 / 0.786026.12 / 0.7868
SwinIR- ℓ\ell26.76 / 0.8156 / 0.27727.39 / 0.819926.62 / 0.8142
SwinIR26.50 / 0.7875 / 0.15127.04 / 0.786726.38 / 0.7878
ESRGAN26.39 / 0.7861 / 0.14526.92 / 0.786726.27 / 0.7858
TABLE I: Quantitative comparisons on DuSR-RealV2-Paired. Bold and underlined indicate best and second-best, respectively. - ℓ\ell denotes training with only reconstruction loss.
MethodFull-ImageCenter-ImageCorner-Image
PSNR ↑\uparrow /SSIM ↑\uparrow /LPIPS ↓\downarrowPSNR/SSIMPSNR/SSIM
RCAN- ℓ\ell26.44 / 0.8676 / 0.14726.91 / 0.870426.33 / 0.8667
SwinIR- ℓ\ell26.14 / 0.8601 / 0.15726.35 / 0.861226.11 / 0.8597
ESRGAN25.78 / 0.8622 / 0.15225.91 / 0.863725.77 / 0.8617
BSRGAN24.77 / 0.8227 / 0.20225.01 / 0.823324.71 / 0.8225
TTSR- ℓ\ell26.48 / 0.8676 / 0.14727.69 / 0.881026.17 / 0.8631
TABLE II: Quantitative comparisons on DuSR-Real.
MethodNRQM ↑\uparrowNIQE ↓\downarrowPI ↓\downarrowBRISQUE ↓\downarrowMUSIQ ↑\uparrowMANIQA ↑\uparrow
RCAN- ℓ\ell4.9295.5905.39143.11553.1490.3295
SwinIR- ℓ\ell4.7636.0665.71442.26251.0870.3154
ESRGAN- ℓ\ell4.6556.1025.77547.83152.1640.3223
BSRGAN- ℓ\ell4.7635.8655.60146.54053.6240.3379
TTSR- ℓ\ell4.6554.8794.96439.63655.5260.3507
MASA- ℓ\ell4.9255.2025.39744.26553.1180.3274
TABLE III: Quantitative comparisons on DuSR-RealV2-Unpaired. Since ground-truth HR images are unavailable, we employ no-reference metrics for objective evaluation.
MethodAdapt.PairedUnpairedParams.Time
PSNR ↑\uparrowSSIM ↑\uparrowPSNR ↑\uparrowSSIM ↑\uparrow(M)(ms)
SPyNet–25.510.876620.600.76421.420
✓\checkmark33.78 (+8.3)(+8.3)0.9470 (+.070)(+.070)23.59 (+3.0)(+3.0)0.8075 (+.043)(+.043)
PWC-Net–28.600.928921.460.79729.416
✓\checkmark33.72 (+5.1)(+5.1)0.9504 (+.022)(+.022)23.47 (+2.0)(+2.0)0.8082 (+.011)(+.011)
RAFT–27.760.881120.840.74405.364
TABLE IV: Comparison of optical flow architectures on DuSR-RealV2 using official pretrained weights and the proposed real–synthetic data fusion strategy. Values in parentheses denote changes relative to the corresponding pretrained models. Latency is measured for estimating one 512×960512\times 960 flow field on an NVIDIA RTX 3090 GPU.
Fig. 8: Visual comparisons on the DuSR-RealV2 dataset. The presented results are obtained with only the reconstruction loss. Zoom in for better observation. The Pseudo ground truth of the unpaired subset is provided for reference only.
VariantsFull-ImageCenter-ImageCorner-Image
No.Ref. WarpingFlow PriorPSNR ↑\uparrow /SSIM ↑\uparrow /LPIPS ↓\downarrowPSNR/SSIMPSNR/SSIM
#1Patch Warpingw/o flow29.07 / 0.8592 / 0.20435.07 / 0.933028.13 / 0.8343
#2Patch WarpingSPyNet29.14 / 0.8652 / 0.19537.22 / 0.957528.09 / 0.8341
#3Patch WarpingDuFlowNet29.22 / 0.8682 / 0.19137.64 / 0.963728.15 / 0.8359
#4Patch + DCNDuFlowNet29.60 / 0.8677 / 0.19437.40 / 0.962828.55 / 0.8356
#5Patch + KF-DCNDuFlowNet29.76 / 0.8711 / 0.19338.43 / 0.968328.68 / 0.8383
TABLE V: Ablation study on DuFlowNet Prior and KF-DCN
VariantsFull-ImageCenter-ImageCorner-Image
PSNR ↑\uparrow /SSIM ↑\uparrow /LPIPS ↓\downarrowPSNR/SSIMPSNR/SSIM
#1:Patch warping29.22 / 0.8682 / 0.19137.64 / 0.963728.15 / 0.8359
#2:KF-DCN warping29.68 / 0.8686 / 0.19737.21 / 0.962228.65 / 0.8370
#3:Region-concat29.34 / 0.8673 / 0.19537.24 / 0.960928.30 / 0.8358
#4:Channel-concat29.59 / 0.8684 / 0.19538.13 / 0.966328.52 / 0.8353
#5:Uniform aggregation29.64 / 0.8692 / 0.19537.42 / 0.962528.61 / 0.8376
TABLE VI: Ablation study on the Complementary Reference Warping
Fig. 9: Visual comparison of optical flow based alignment results on the DuSR-RealV2-Unpaired subset. The artifacts in the figure indicate the displacement differences between the aligned results and the LR images.
Fig. 10: Visualization of reference feature aggregation results. Green and red boxes indicate center and corner regions, respectively.
Fig. 11: Failure case under extremely large parallax. The white dashed box indicates the center region, and the yellow box marks the enlarged area. The color artifacts in the Ref image visualize the large LR–Ref displacement. Under such extreme parallax, DuSRFlow still exhibits residual fusion artifacts, although fewer than KeDuSR.
MethodFull-ImageCenter-ImageCorner-Image
PSNR ↑\uparrow /SSIM ↑\uparrow /LPIPS ↓\downarrowPSNR/SSIMPSNR/SSIM
RCAN- ℓ\ell25.96 / 0.8033 / 0.23425.69 / 0.793726.12 / 0.8065
SwinIR- ℓ\ell25.78 / 0.7982 / 0.24625.50 / 0.788525.94 / 0.8015
TTSR- ℓ\ell25.92 / 0.8017 / 0.23525.94 / 0.796225.98 / 0.8036
MASA- ℓ\ell25.95 / 0.7989 / 0.23925.81 / 0.789926.07 / 0.8020
DATSR- ℓ\ell25.81 / 0.7975 / 0.24225.58 / 0.788225.95 / 0.8007
TABLE S1: Quantitative comparisons on RealMCVSR-Real.
 Full-ImageCenter-ImageCorner-Image
MethodPSNR ↑\uparrow /SSIM ↑\uparrow /LPIPS ↓\downarrowPSNR/SSIMPSNR/SSIM
RCAN- ℓ\ell25.67 / 0.8049 / 0.30826.65 / 0.815825.45 / 0.8012
SwinIR- ℓ\ell25.32 / 0.8007 / 0.31525.81 / 0.807325.22 / 0.7985
TTSR- ℓ\ell25.83 / 0.8044 / 0.31126.75 / 0.818825.62 / 0.7996
MASA- ℓ\ell25.78 / 0.8030 / 0.30326.70 / 0.815525.58 / 0.7988
DCSR- ℓ\ell26.02 / 0.8123 / 0.29328.37 / 0.844025.51 / 0.8016
TABLE S2: Quantitative comparisons on CameraFusion-Real.
MethodDuSR-RealRealMCVSR-RealCameraFusion-Real
PSNR ↑\uparrow /SSIM ↑\uparrow /LPIPS ↓\downarrowPSNR ↑\uparrow /SSIM ↑\uparrow /LPIPS ↓\downarrowPSNR ↑\uparrow /SSIM ↑\uparrow /LPIPS ↓\downarrow
TTSR- ℓ\ell24.59 / 0.8374 / 0.17523.15 / 0.7513 / 0.27325.58 / 0.8018 / 0.279
MASA- ℓ\ell24.84 / 0.8270 / 0.19824.40 / 0.7575 / 0.28525.43 / 0.7947 / 0.289
DATSR- ℓ\ell24.58 / 0.8278 / 0.19624.01 / 0.7610 / 0.27125.17 / 0.7917 / 0.288
DCSR- ℓ\ell25.30 / 0.8497 / 0.16924.69 / 0.7807 / 0.25125.56 / 0.8054 / 0.276
SelfDZSR- ℓ\ell25.79 / 0.8402 / 0.18824.82 / 0.7703 / 0.27426.35 / 0.8057 / 0.290
TABLE S3: Generalization evaluation of models trained on DuSR-RealV2.
DatasetMetricsw/o parallax sim.Synthetic-onlyOurs
PairedPSNR ↑\uparrow33.7231.1533.74
SSIM ↑\uparrow0.95070.90500.9510
UnpairedPSNR ↑\uparrow22.4322.8323.29
SSIM ↑\uparrow0.77880.75260.8164
TABLE S4: Ablation study of our real–synthetic data fusion strategy.
Fig. S1: Visual comparison of alignment results. The red and green boxes highlight large-parallax foreground and background regions, respectively.
VariantPeak Memory (MiB)
Patch warping only2371.09
KF-DCN warping only3502.75
Ours3600.75
TABLE S5: Inference-time peak GPU memory of different warping configurations.
MethodSISRRefSR
RCANSwinIRESRGANBSRGANTTSRMASADATSRDCSRSelfDZSRZeDuSRKeDuSROurs
Latency (s)0.692.850.080.457.511.529.350.840.171800.510.47
Params (M)5.6311.7516.7016.616.254.0210.513.190.521.495.635.02
FLOPs (T)6.150.411.806.772.922.345.812.081.14——2.462.76
TABLE S6: FLOPs and latency are measured for generating an HR output of size 1792×8961792\times 896 on a single NVIDIA RTX 3090 GPU.
Fig. S2: Visual comparisons on RealMCVSR-Real, DuSR-Real, and CameraFusion-Real datasets. Zoom in for better observation.
Fig. S3: Visual comparisons of SR results on the DuSR-RealV2 dataset. The upper part shows results on the paired subset, and the lower part shows results on the unpaired subset. The presented results are obtained with only the reconstruction loss. For the unpaired subset, the pseudo ground truth is provided for reference only. Zoom in for better observation.
More in this topic

Reference-Based Image Super-Resolution

See all 10 papers

Explore similar work

CardsList
May 12, 2026

Fast Image Super-Resolution via Consistency Rectified Flow

Diffusion models (DMs) have demonstrated remarkable success in real-world image super-resolution (SR), yet their reliance on time-consuming multi-step sampling largely hinders their practical applications. While recent efforts have introduced few- or single-step solutions, existing methods either inefficiently model the process from noisy input or fail to fully exploit iterative generative priors, compromising the fidelity and quality of the reconstructed images. To address this issue, we propose FlowSR, a novel approach that reformulates the SR problem as a rectified flow from low-resolution (LR) to high-resolution (HR) images. Our method leverages an improved consistency learning strategy to enable high-quality SR in a single step. Specifically, we refine the original consistency distillation process by incorporating HR regularization, ensuring that the learned SR flow not only enforces self-consistency but also converges precisely to the ground-truth HR target. Furthermore, we introduce a fast-slow scheduling strategy, where adjacent timesteps for consistency learning are sampled from two distinct schedulers: a fast scheduler with fewer timesteps to improve efficiency, and a slow scheduler with more timesteps to capture fine-grained texture details. Extensive experiments demonstrate that FlowSR achieves outstanding performance in both efficiency and image quality. Code: this https URL.

Jiaqi Xu, Wenbo Li, Haoze Sun +8

The Chinese University of Hong Kong · Huawei Noah’s Ark Lab · HKUST (GZ) +1

Rectified FlowConsistency DistillationDiffusion Model AccelerationImage Super-Resolution
AskPDF
May 18, 2026

Learning to Balance: Decoupled Siamese Diffusion Transformer for Reference-Based Remote Sensing Image Super-Resolution

Diffusion-based methods demonstrate significant potential for remote sensing image super-resolution at large scaling factors, particularly in reference-based super-resolution (RefSR), where high-resolution reference images provide critical fine-grained texture priors. However, existing methods often suffer from a trade-off between over-reliance on reference information, which leads to texture artifacts, and under-utilization of such information, which results in insufficient detail recovery. To address these issues, we propose DS-DiT, a Decoupled Siamese Diffusion Transformer that decouples the interaction between low-resolution (LR) and reference (Ref) conditions within the attention mechanism. By allowing LR structural priors and Ref texture information to independently interact with the noisy latent, the framework effectively mitigates competition between the two conditional sources. To further compensate for the limited local modeling ability of global attention, we introduce a Patch-Level Weighting (PLW) module that adaptively modulates the fusion of conditional sources. In addition, the siamese architecture enables an inference-time autoguidance strategy that exploits the prediction discrepancy between strong and weak Ref conditions to improve generation quality without additional training. Experimental results across multiple datasets and scaling factors show that DS-DiT outperforms existing methods in both quantitative metrics and visual fidelity.

Bin Luo, Runmin Dong, Zhaoyang Luo +4

Tsinghua Shenzhen International Graduate School, Shenzhen, China · Sun Yat-sen University, Zhuhai, China · Tsinghua University, Beijing, China +1

Diffusion Models for Image RestorationDiffusion TransformerDiffusion GuidanceRemote Sensing Image Super-ResolutionReference-Based Image Super-ResolutionImage Super-Resolution
AskPDF
Jul 14, 2026

RFMSR: Residual Flow Matching for Image Super-Resolution

Image super-resolution (ISR) has witnessed remarkable progress with diffusion models and flow matching. The dominant text-to-image (T2I) based approaches leverage large-scale foundation models as generative priors, achieving impressive perceptual quality but at the cost of massive model sizes and prohibitive training expenses. Recent flow-matching-based vision-only approaches have made significant strides; however, they adopt standard flow formulations that transport from a pure Gaussian prior to the data distribution, discarding the rich structural information already present in the low-quality (LQ) input. Furthermore, existing single-step acceleration techniques often forfeit the model's multi-step inference capability. In this paper, we propose Residual Flow Matching for Image Super-Resolution (RFMSR), a vision-only framework that centers the source distribution at the LQ latent, reducing transport distance and preserving structural priors throughout the flow trajectory. We further introduce a two-phase training strategy: Phase I pretrains the velocity field via conditional flow matching, while Phase II applies end-to-end supervision to the single-step prediction while retaining the velocity loss across all timesteps, achieving high-quality single-step generation without sacrificing multi-step refinement. Extensive experiments demonstrate that RFMSR achieves comparable or even superior perceptual quality compared to state-of-the-art (SOTA) methods. The source code is available at https://github.com/Faze-Hsw/RFMSR.

Shuwei Huang, Tianyao Luo, Jicheng Liu +2

Huazhong University of Science and Technology · Wuhan University

Flow MatchingImage Super-Resolution
AskPDF