Stereo Super-Resolution (SSR)

• Stereo Super-Resolution (SSR) is a technique that reconstructs high-resolution stereo pairs by jointly leveraging intra-view structures and cross-view correspondences.
• It employs diverse methods such as CNN-based parallax attention, Transformer architectures, and diffusion models to enhance image fidelity and maintain geometric consistency.
• Applications span medical endoscopy, stereoscopic video processing, and 3D reconstruction, addressing challenges like occlusion, disparity distortion, and real-world degradation.

Stereo Super-Resolution (SSR) is the task of reconstructing a high-resolution stereo pair from a low-resolution stereo pair by exploiting both intra-view image structure and cross-view correspondence. In the binocular case, SSR learns a mapping from left and right low-resolution inputs to left and right super-resolved outputs while preserving the disparity structure that underlies stereo perception and 3D geometry. A representative formulation is

$$f_\theta:(I_L^{LR}, I_R^{LR}) \mapsto (I_L^{SR}, I_R^{SR}),$$

with the additional requirement that the two outputs remain mutually consistent under stereo correspondence rather than being treated as two unrelated single-image SR problems (Zhang et al., 2020). In the broader SR taxonomy, SSR is usually distinguished from SISR, VSR, and LFSR by its explicit use of cross-view disparity correlation, with inputs often written as $$X_{lr}^i \in \mathbb{R}^{M \times H \times W \times C}$$ and outputs as $$X_{hr}^i \in \mathbb{R}^{M \times sH \times sW \times C}$$, where $M$ denotes the number of views (Zhang et al., 19 Sep 2025).

1. Geometric formulation and problem characteristics

SSR differs from single-image SR because the two views are related by epipolar geometry, horizontal disparity, and partial occlusion. Rectified stereo pairs satisfy the constraint that corresponding points lie on the same row, which makes cross-view fusion fundamentally a correspondence problem along the width dimension. The second view contributes redundant and complementary observations: the same scene structure appears from a slightly different viewpoint, some details occluded in one view may be visible in the other, and repeated textures can be disambiguated by binocular agreement (Wang et al., 2020).

This additional information also introduces extra failure modes. A stereo SR model must not only sharpen edges and restore fine textures, but also avoid left-right inconsistencies, disparity distortion, and occlusion-induced artifacts. In endoscopic stereo, these issues are compounded by small baselines, deformation, specular highlights, non-uniform illumination, smoke, and imperfect rectification (Zhang et al., 2020). In stereoscopic video, temporal-consistency becomes an additional requirement, because stereo-consistency and temporal-consistency jointly affect 3D comfort and “3D fatigue” (Imani et al., 2022).

The survey literature therefore treats SSR as a geometry-constrained restoration problem rather than a simple two-image extension of SISR. The central design question is how to exploit cross-view information strongly enough to improve SR, yet selectively enough to avoid transferring mismatched content from the auxiliary view (Zhang et al., 19 Sep 2025).

2. Architectural families and representative models

A common taxonomy divides SSR methods into CNN-based regression models, Transformer-based regression models, and generative models including GAN-based and diffusion-based approaches (Zhang et al., 19 Sep 2025). Within CNN-based regression, the dominant pattern is a symmetric two-branch network with shared feature extractors and an explicit cross-view interaction module.

Parallax-attention architectures were a decisive step in this line. "Symmetric Parallax Attention for Stereo Image Super-Resolution" introduced iPASSR, which uses a Siamese network with a symmetric bi-directional parallax attention module (biPAM), inline occlusion handling, and bilateral stereo-consistency losses. The attention maps $\mathbf{M}_{R\rightarrow L}$ and $\mathbf{M}_{L\rightarrow R}$ have shape $H\times W\times W$, so each pixel attends along its epipolar line, and cross-view converted features are formed by batch-wise matrix multiplication with the opposite-view feature map (Wang et al., 2020). In the medical domain, "Stereo Endoscopic Image Super-Resolution Using Disparity-Constrained Parallel Attention" generalizes this idea with an Atrous Parallax-Attention Module (APAM) that uses multi-line similar features around the epipolar line and enforces disparity consistency between LR and SR representations (Zhang et al., 2020).

A second family makes disparity itself an explicit partner of SR. "Feedback Network for Mutually Boosted Stereo Image Super-Resolution and Disparity Estimation" proposes SSRDE-FNet, which jointly performs stereo SR and disparity estimation in a unified framework. It first exploits cross-view information in LR space, then estimates HR disparity from SR-produced HR features, uses HR disparity to aggregate HR features for finer SR, and finally feeds HR disparity information back to previous layers through an HR Disparity Information Feedback mechanism (Dai et al., 2021). This design treats SR and disparity as mutually reinforcing tasks rather than separate modules.

A third family emphasizes simplicity and efficient baselines. "NAFSSR: Stereo Image Super-Resolution Using NAFNet" adopts NAFNet as a shared intra-view backbone and adds a Stereo Cross-Attention Module (SCAM) after NAFBlocks, producing a deliberately simple yet strong stereo SR baseline. It reports that NAFSSR outperforms previous state-of-the-art methods on KITTI 2012, KITTI 2015, Middlebury, and Flickr1024, and that it won 1st place in the NTIRE 2022 Stereo Image Super-resolution Challenge (Chu et al., 2022). "NAFRSSR: a Lightweight Recursive Network for Efficient Stereo Image Super-Resolution" continues this efficiency-oriented line with NAFGCBlocks, DSSCAM, recursion, and a trainable edge detector; the smallest variant, NAFRSSR-M, has 0.28M parameters and 50 ms inference time while reaching 24.657 dB / 0.7622 average PSNR/SSIM on the benchmark datasets (Chen et al., 2024).

More recent CNN models strengthen intra-view modeling and selective fusion rather than merely enlarging stereo attention. "Cross-View Hierarchy Network for Stereo Image Super-Resolution" emphasizes that many prior methods over-prioritized cross-view fusion and underused intra-view hierarchy. Its CHIMB and CVIM pair large-kernel convolution attention and channel attention for intra-view modeling with depth-wise-convolution-enhanced cross-view attention (Zou et al., 2023). "Learning Accurate and Enriched Features for Stereo Image Super-Resolution" proposes MSSFNet, combining a Mixed-Scale Block, a Selective Fusion Attention Module, and Fast Fourier Convolution Blocks to preserve spatial detail, enrich context, and filter irrelevant cross-view similarities before attention (Gao et al., 2024).

Transformer-based SSR methods explicitly separate or interleave cross-view and intra-view attention. "PFT-SSR: Parallax Fusion Transformer for Stereo Image Super-Resolution" uses a Swin-based backbone with a Parallax Fusion Transformer consisting of a Cross-view Fusion Transformer (CVFT) and an Intra-view Refinement Transformer (IVRT). The design is explicitly motivated by the mismatch between CNN-era stereo fusion modules and Transformer backbones, and by the need to handle varying parallax through alternating cross-view interaction and local refinement (Guo et al., 2023).

The most recent extensions move beyond fixed-scale convolutional upsampling and beyond purely deterministic regression. "StereoINR: Cross-View Geometry Consistent Stereo Super Resolution with Implicit Neural Representation" models stereo images as continuous implicit functions, combines a HAT-based encoder with disparity-guided warping and cross-attention, and replaces fixed pixel-shuffle upsampling with coordinate-based arbitrary-scale reconstruction (Liu et al., 7 May 2025). "DIffSteISR: Harnessing Diffusion Prior for Superior Real-world Stereo Image Super-Resolution" introduces a diffusion-based real-world SSR framework with time-aware stereo cross attention, stereo ControlNet conditioning, and stereo semantic extraction, targeting texture realism and cross-view semantic consistency under real degradations (Zhou et al., 2024).

3. Correspondence modeling, losses, and optimization objectives

The core mathematical machinery of SSR is cross-view correspondence. In parallax-attention models, a left-view feature attends over candidate positions in the right view along the same row, producing a distribution over disparities. In iPASSR, cross-view converted features are written as

$$\mathbf{F'}_{R\rightarrow L} = \mathbf{M}_{R\rightarrow L}\otimes \mathbf{F}_R,\qquad \mathbf{F'}_{L\rightarrow R} = \mathbf{M}_{L\rightarrow R}\otimes \mathbf{F}_L,$$

and are then fused with intra-view features through validity masks inferred from cycle-consistency of attention maps (Wang et al., 2020). The survey formalizes the same pattern as a generic SSR loss framework built on parallax attention maps $\mathbf{M}_{left\rightarrow right}$ and $\mathbf{M}_{right\rightarrow left}$, together with photometric consistency, smoothness, and cycle-consistency losses (Zhang et al., 19 Sep 2025).

A common reconstruction objective is an $$X_{lr}^i \in \mathbb{R}^{M \times H \times W \times C}$$0 or $$X_{lr}^i \in \mathbb{R}^{M \times H \times W \times C}$$1 loss over both views. NAFSSR uses the simple stereo $$X_{lr}^i \in \mathbb{R}^{M \times H \times W \times C}$$2 objective

$$X_{lr}^i \in \mathbb{R}^{M \times H \times W \times C}$$3

whereas DCSSRnet uses an MSE reconstruction term plus a disparity-consistency penalty that aligns LR and SR disparity masks: $$X_{lr}^i \in \mathbb{R}^{M \times H \times W \times C}$$4 The latter explicitly ties the stereo geometry of the SR output to the input stereo pair (Chu et al., 2022, Zhang et al., 2020).

Disparity-aware models add dedicated geometry objectives. SSRDE-FNet combines an SR loss, a BiPAM loss, and a disparity loss: $$X_{lr}^i \in \mathbb{R}^{M \times H \times W \times C}$$5 with $$X_{lr}^i \in \mathbb{R}^{M \times H \times W \times C}$$6. Its disparity module uses HR features generated by SR to estimate HR disparity through a coarse-to-fine cost-volume refinement, then feeds HR disparity back into earlier layers (Dai et al., 2021). This establishes a stronger form of geometric coupling than attention-only methods.

Perception-oriented SSR replaces generic perceptual supervision with stereo-specific quality supervision. "Perception-Oriented Stereo Image Super-Resolution" introduces StereoSRQA and trains the SR network with a pixel loss plus image-level and feature-level IQP losses in the learned StereoSRQA feature space: $$X_{lr}^i \in \mathbb{R}^{M \times H \times W \times C}$$7 This design is motivated by the observation that VGG- or GAN-based perceptual objectives can introduce stereo-inconsistent artifacts even when single-view sharpness improves (Ma et al., 2022).

Diffusion-based real-world SSR changes the optimization regime again. DiffSteISR uses a diffusion noise-prediction loss over stereo latents conditioned on LR stereo latents and semantic prompts: $$X_{lr}^i \in \mathbb{R}^{M \times H \times W \times C}$$8 Here cross-view consistency is not imposed by a separate disparity loss, but by stereo-conditioned denoising, time-aware stereo cross attention, and stereo-aware control features (Zhou et al., 2024).

4. Datasets, benchmarks, and evaluation protocols

The modern empirical foundation of SSR is Flickr1024, a dataset of 1024 high-quality stereo pairs collected from Flickr and processed for stereo SR. The dataset is split into 800 training pairs, 112 validation pairs, and 112 test pairs. Its processing includes reordering cross-eye stereo into standard left-right pairs, discarding vertically misaligned images, cropping to remove margins, and enforcing a disparity configuration in which close-shot images satisfy $$X_{lr}^i \in \mathbb{R}^{M \times H \times W \times C}$$9 pixels (Wang et al., 2019). Cross-dataset experiments in the Flickr1024 paper show that training StereoSR and PASSRnet on Flickr1024 yields the best PSNR and SSIM on KITTI2015, Middlebury, Flickr1024, and ETH3D, and that the larger, more diverse dataset mitigates over-fitting (Wang et al., 2019).

Beyond Flickr1024, the standard SSR benchmarks are KITTI 2012, KITTI 2015, Middlebury, and sometimes Tsukuba. Many later methods train on 800 Flickr1024 pairs plus 60 Middlebury pairs and evaluate on KITTI 2012, KITTI 2015, Middlebury, and Flickr1024 at $$X_{hr}^i \in \mathbb{R}^{M \times sH \times sW \times C}$$0 and $$X_{hr}^i \in \mathbb{R}^{M \times sH \times sW \times C}$$1 scale factors (Guo et al., 2023). MSSFNet follows this protocol and also reports two evaluation styles: Left-only PSNR/SSIM with a 64-pixel left boundary crop, and stereo average $$X_{hr}^i \in \mathbb{R}^{M \times sH \times sW \times C}$$2 without boundary cropping (Gao et al., 2024).

PSNR and SSIM remain the default distortion metrics, but recent work broadens evaluation to stereo-specific and perception-oriented criteria. The survey writes PSNR as

$$X_{hr}^i \in \mathbb{R}^{M \times sH \times sW \times C}$$3

and treats SSIM as a standard structural fidelity measure (Zhang et al., 19 Sep 2025). PSSR supplements them with NIQE, PI, BRISQUE, SSEQ, CNNIQA, SFA, StereoQA, and StereoSRQA, arguing that PSNR and SSIM alone do not track binocular perceptual quality (Ma et al., 2022). DiffSteISR further uses LPIPS, DISTS, FID, MADE, NIQE, MANIQA, MUSIQ, and CLIPIQA; MADE is defined as a mean absolute disparity error between disparity maps of the super-resolved stereo pair and the ground-truth pair, so it directly measures geometric fidelity (Zhou et al., 2024).

These choices matter because recent models are no longer optimized for a single notion of quality. GAN- and diffusion-based methods often trade some PSNR for perceptual realism, while geometry-aware methods may improve disparity consistency even when the pixel-wise gain is modest. The survey therefore treats SSR evaluation as intrinsically multi-criteria: distortion, perceptual fidelity, and stereo consistency (Zhang et al., 19 Sep 2025).

5. Domain-specific deployments and downstream uses

One of the clearest domain-specific instantiations of SSR is stereo endoscopy. DCSSRnet was developed for stereo laparoscopic images captured by the stereo camera of the da Vinci surgical system, using 4560 HR stereo pairs of size $$X_{hr}^i \in \mathbb{R}^{M \times sH \times sW \times C}$$4. It addresses limited sensor size, narrow field of view, fine tissue structure, deformation, specular highlights, blood, smoke, and small baselines by combining a residual ASPP backbone with APAM and disparity-consistency losses, and it reports improvements over existing SR methods on laparoscopic data (Zhang et al., 2020). This medical setting makes explicit a general SSR requirement: SR should enhance resolution without degrading depth perception.

SEGSRNet extends this logic to a task pipeline in which stereo super-resolution precedes surgical instrument segmentation. Its SR module uses CCSB, ASPP, RDBs, and a biPAM-style cross-view interaction module, and its segmentation stage uses an SPP-LinkNet-34 variant. The paper reports that SEGSRNet reaches 42.41 / 0.9879 on MICCAI 2018 at $$X_{hr}^i \in \mathbb{R}^{M \times sH \times sW \times C}$$5 and 36.01 / 0.9768 at $$X_{hr}^i \in \mathbb{R}^{M \times sH \times sW \times C}$$6, while the segmentation component achieves 83.65 ± 16.47 IoU and 89.80 ± 13.99 Dice on binary segmentation in EndoVis 2017 (Hayat et al., 2024). The broader implication is that stereo-aware SR can be task-oriented even when training remains stage-wise.

Stereo video SR generalizes SSR into a spatio-temporal setting. "A New Dataset and Transformer for Stereoscopic Video Super-Resolution" introduces SVSR-Set, a dataset of 71 full HD stereo videos captured using a professional stereo camera, and proposes Trans-SVSR with a spatio-temporal convolutional self-attention layer, an optical flow-based feed-forward layer, and PAM-based stereo fusion. The method is explicitly motivated by the need to preserve both stereo-consistency and temporal-consistency, and it reports the lowest disparity end-point error, 0.5031, among the compared methods on SVSR-Set (Imani et al., 2022).

SSR is also increasingly evaluated by downstream geometry. iPASSR reports that when GwcNet is run on $$X_{hr}^i \in \mathbb{R}^{M \times sH \times sW \times C}$$7 SR outputs on SceneFlow, the end-point error improves from 1.196 for bicubic-upsampled LR to 0.949 for iPASSR, compared with 0.667 on HR ground truth (Wang et al., 2020). At the system level, "Improving Multi-View Stereo via Super-Resolution" shows that inserting super-resolution before MVS generally improves reconstructed 3D models, especially completeness and especially on textured scenes, even when the SR module is monocular rather than stereo-aware (Lomurno et al., 2021). This does not make the paper an SSR method, but it clarifies why geometry-consistent SR is operationally important beyond image quality.

6. Real-world, perception-oriented, efficient, and arbitrary-scale directions

Recent SSR work has moved toward four partially overlapping goals: perceptual realism, real-world degradation robustness, efficiency, and scale flexibility. PSSR is the first perception-oriented StereoSR approach built around a dedicated StereoSRQA model and a StereoSRQA database. Mid3D_QA contains 3366 stereo SR pairs generated from 66 Middlebury HR stereo pairs, and FRStereoSRQA achieves RMSE 0.3505, SROCC 0.9910, PLCC 0.9918, and KROCC 0.9170 on that benchmark (Ma et al., 2022). This line reframes SSR evaluation as an explicitly learned binocular quality problem.

DiffSteISR addresses real-world stereo image super-resolution by importing Stable Diffusion 2.0 as a texture prior and constraining it with stereo-specific mechanisms: TASCATA, SOA ControlNet, and SSE. On StereoWeb20, which has no ground truth, it reports the best MANIQA, MUSIQ, and CLIPIQA scores among the compared methods, while on synthetic Flickr1024RS it substantially improves MADE over other diffusion-based baselines (Zhou et al., 2024). A plausible implication is that diffusion-based SSR will remain competitive primarily when semantic realism and real degradations are dominant concerns, provided stereo consistency is explicitly modeled.

Efficiency remains a parallel research axis rather than a secondary concern. NAFRSSR shows that careful redesign of attention and convolution blocks can improve the accuracy-efficiency trade-off beyond NAFSSR and SwinFIR, and MSSFNet reports state-of-the-art PSNR with fewer FLOPs than SwinFSR, Steformer, and many others, including up to 87% fewer parameters than SwinFSR (Chen et al., 2024, Gao et al., 2024). CVHSSR similarly reports that CVHSSR-S matches or exceeds NAFSSR-B while using roughly 70% fewer parameters, emphasizing that stronger intra-view modeling can replace some of the complexity of heavy stereo fusion (Zou et al., 2023).

Arbitrary-scale SSR is emerging as a distinct subproblem. StereoINR replaces fixed-scale pixel-shuffle upsampling with continuous implicit neural representations conditioned on disparity-guided cross-view fusion, allowing scales such as $$X_{hr}^i \in \mathbb{R}^{M \times sH \times sW \times C}$$8 and larger without architectural change. It reports that StereoINR outperforms arbitrary-scale single-image INR baselines on Middlebury and matches state-of-the-art fixed-scale SSR methods within training-distribution scales while improving stereo consistency scores (Liu et al., 7 May 2025).

The current survey identifies the resulting open directions with unusual clarity: real-world blind SSR under unknown view-dependent degradations, joint SSR and depth estimation, arbitrary-scale SSR, larger and more diverse stereo datasets, and unsupervised or self-supervised SSR that exploits left-right photometric constancy, smoothness, and cycle-consistency without requiring HR supervision (Zhang et al., 19 Sep 2025). This suggests that the field is converging on an overview of geometry-aware correspondence, stronger generative priors, and task-aware evaluation, rather than on any single backbone family.