PSISRD_VAL125

Public Single Image Super-Resolution Dataset — Validation Set (125 Images)
An open, CC0-licensed validation dataset for benchmarking and evaluating Single Image Super-Resolution (SISR) models.
Freely available for research, development, and commercial use — no attribution required.

🖼️ Example Subset

Below is an example showing the first 36 images from the PSISRD_VAL125 dataset.
This gives a visual impression of the content and quality of the dataset.

---

📦 Dataset Overview

This dataset contains 125 high-resolution (HR) images curated for validating and benchmarking SISR models.
It is primarily intended for computing evaluation metrics during model training or cross-model comparison.

• All images originate from the UHD-IQA dataset (validation & test subsets) under the CC0 1.0 Universal license.
• Images were downscaled (and cropped where necessary) to 512×512 using the DPID method via pepedpid with λ ≈ 1.0 to preserve fine detail and sharpness.
• The dataset includes corresponding x2 and x4 low-resolution subsets for use as inputs to SISR models.

---

🧪 Validation Purpose

This dataset serves as a public benchmark for SISR model validation.
It has been used for internal validation of the 2xPublic_realplksr_dysample_layernorm model family — see
👉 github.com/Phhofm/models

---

🔍 Quality Metric Criteria

Each image was evaluated using a set of objective quality metrics to ensure high perceptual fidelity and naturalness.
Images failing to meet any single threshold were excluded, except for a few minor edge cases described below.

# Default quality thresholds – tuned for SISR validation
DEFAULT_THRESHOLDS = {
    'brisque': 20,          # Pixel-level artifacts (<=)
    'hyperiqa': 0.7,        # Perceptual quality (>=)
    'maniqa': 0.6,          # Holistic quality (>=)
    'clipiqa_plus': 0.38,   # Semantic plausibility (>=)
    'topiq_nr': 0.6,        # Naturalness (>=)
    'musiq': 65,            # Multi-scale quality (>=)
    'tres': 80,             # Transformer-based quality (>=)
    'arniqa': 0.65,         # Agnostic quality (>=)
    'blockiness': 10,       # Compression artifacts (<)
    'complexity': 0.2       # Texture detail (>=)
}

The evaluation script is optimized to compute faster metrics first, quickly rejecting low-quality samples.

---

📈 Dataset Composition & Deviations

• Total images: 125
• Strictly pass all thresholds: ~121
• Minor deviations: ~4 (included for diversity and edge-case coverage)

For transparency, the full IQA metric scores for every image are provided in the included scores.csv file, along with pass/fail flags for each metric.

---

⚠️ Notes on Deviations

A few samples slightly deviate from the strict metric thresholds but remain visually excellent and valuable for validation:

These outliers provide useful edge cases for SISR model validation.

---

📄 License

All images are distributed under the Creative Commons Zero (CC0 1.0 Universal) license. ✅ Free for any purpose — research, commercial, or derivative use — no attribution required.

---

🔗 Related Resources

• Training Dataset: PSISRD — currently 13,761 images (512×512) curated and filtered the same way as PSISRD_VAL125. Fully open under CC0, and can be used freely for training SISR models.
• Models: 2xPublic_realplksr_dysample_layernorm — trained on PSISRD, validated on PSISRD_VAL125 → GitHub repo
• Downscaling Method: pepedpid (DPID implementation)