SRMORSS: Super-Resolution Model-Oriented Realistic Self-Supervision Dataset

Purpose of the Dataset

SRMORSS is designed to support the image super-resolution (SR) research community by providing a large and comprehensive dataset composed of images reconstructed by a diverse set of modern and classical SR models from different methodological categories. These reconstructions are generated from a substantial collection of realistic low-resolution (LR) input images.

The primary objective of this dataset is to facilitate the training of modern machine learning and deep learning models for constructing a meaningful SR distortion manifold. This manifold can be interpreted as a latent space that captures informative representations of quality degradation patterns specific to the SR domain. By incorporating outputs from multiple SR models, called model-oriented approach, the dataset enables the learning of a model-agnostic latent representation of SR-induced distortions. Such a representation is particularly valuable for identifying and characterizing degradation patterns produced by previously unseen or novel SR methods, thereby supporting a wide range of downstream tasks.

A key distinguishing feature of SRMORSS is its focus on real-world scenarios. Unlike many existing SR datasets that rely on synthetically generated LR images—typically produced by downsampling high-resolution (HR) images with predefined degradations such as blurring—SRMORSS is constructed using authentic LR inputs. This design choice better reflects practical SR deployment conditions and captures the complexity and variability of real-world degradations.

The dataset is particularly well suited for super-resolution image quality assessment (SR-IQA) metric learning. By including reconstructions from a broad spectrum of SR algorithms applied to numerous real LR images, SRMORSS addresses a significant gap in existing SR-IQA datasets, which often lack diversity in degradation types or rely on synthetic data generation pipelines. Although subjective quality scores are not provided for the reconstructed images, the dataset is ideally suited for use in unsupervised pretraining or as a pretext task within self-supervised learning (SSL) frameworks for SR-IQA. The usefulness of this dataset has been demonstrated in our work [1]: “Self-Supervised Image Super-Resolution Quality Assessment based on Content-Free Multi-Model Oriented Representation Learning”. 

Beyond SR-IQA, SRMORSS is broadly applicable to any research task that requires a comprehensive understanding of SR-related distortions and their characteristics.

Construction Details 

The source low- and high-resolution images used in this dataset are derived from the RealSR dataset [2]. Building upon this source, we construct our new large-scale and unsupervised dataset by generating super-resolved images using a diverse set of SR models and scaling factors. Specifically, we selected 482 reference images from RealSR that provide all three scaling factors (x2, x3, and x4). Using these references, we generated a total of 15,374 super-resolved images produced by 13 different SR models, by applying those models on the selected LR images.
This design enables the collection of rich, unlabeled distortions across algorithms and scales, making the dataset particularly well suited for self-supervised learning (SSL) and representation learning in realistic SR scenarios.

Dataset Structure Description

The dataset is organized into three primary directories: HR, LR, and SR, corresponding to high-resolution ground truth images, low-resolution inputs, and super-resolved outputs, respectively.

The LR directory is further subdivided according to the downsampling factor, with separate folders for each scale (x2, x3, and x4). Each LR image follows a consistent naming convention that preserves the identity of the original scene and the associated scale factor.

The SR directory is organized hierarchically based on the super-resolution paradigm, the specific SR algorithm, and the upscaling factor. At the top level, SR methods are grouped into broad categories, including Interpolation, CNN-based, GAN-based, Transformer-based, Diffusion-based, GNN, INR, and Lightweight approaches. Within each category, subfolders correspond to individual algorithms, which are further divided into scale-specific directories (x2, x3, x4).

All images adhere to a unified naming convention of the form:

<SceneID>_x<Scale>_<Method>[_Variant].png

where <SceneID> denotes the source image, <Scale> indicates the magnification factor, <Method> specifies the SR algorithm, and optional suffixes capture model variants when applicable.

SR/
├── CNN
│   ├── SRCNN
│   │   ├── x2
│   │   ├── x3
│   │   └── x4
│   ├── USRNet
│   │   ├── x2
│   │   ├── x3
│   │   └── x4
│   └── VDSR
│       ├── x2
│       ├── x3
│       └── x4
├── Diffusion
│   └── InvSR
│       └── x4
├── GAN
│   ├── BSRGAN
│   │   ├── x2
│   │   └── x4
│   ├── Real-ESRGAN
│   │   ├── x2
│   │   ├── x3
│   │   └── x4
│   └── SRGAN
│       ├── x2
│       └── x4
├── GNN
│   └── IPG
│       └── x4
├── INR
│   └── HIIF
│       ├── x2
│       ├── x3
│       └── x4
├── Interpolation
│   └── Bicubic
│       ├── x2
│       ├── x3
│       └── x4
├── Lightweight
│   ├── CATANet
│   │   └── x4
│   └── SeemoRe
│       ├── x2
│       ├── x3
│       └── x4
└── Transformer
    └── SwinIR
        ├── x2
        ├── x3
        └── x4

This structured design enables systematic benchmarking across resolutions, methods, and SR paradigms, while ensuring clarity, scalability, and reproducibility.

[1] Kian Majlessi, Amir Masoud Soltani, Mohammad Ebrahim Mahdavi, Aurelien Gourrier, and Peyman Adibi, "Self-Supervised Image Super-Resolution Quality Assessment based on Content-Free Multi-Model Oriented Representation Learning", arXiv, 2026, [online available] https://arxiv.org/abs/2602.10744, [date of access] Feb. 2026.

[2] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang, "Toward Real-World Single Image Super-Resolution: A New Benchmark and a New Model", Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 3086-3095, 2019.