README.md

metadata

license: other
pretty_name: 'DEAR: Dataset for Evaluating the Aesthetics of Rendering (100-scene sample)'
tags:
  - computer-vision
  - image-quality-assessment
  - aesthetics
  - rendering
  - human-preference
task_categories:
  - image-classification
  - image-to-image
  - other
size_categories:
  - 1K-10K

REPID: Rendering Evaluation of Photographic Image Dataset

REPID (officially introduced as the Rendering Evaluation of Photographic Image Dataset) is a large-scale benchmark designed for Image Rendering Quality Assessment (IRQA).

Unlike traditional Image Quality Assessment (IQA) which focuses on technical degradations like noise or blur, REPID aims to model subjective human aesthetic preferences for different rendering styles of the same scene.

Dataset Overview

Built upon the MIT-Adobe FiveK dataset, REPID provides a massive collection of pairwise human preference annotations for professional and automated renderings.

• Scenes: 5,000 high-resolution RAW photographs.
• Total Images: 30,000 unique renderings (6 per scene).
• Total Votes: Over 2.5 million unique votes collected via crowdsourcing.
• Annotators: 13,648 unique evaluators, with each image pair receiving at least 25 individual votes.
• Comparison Task: For each of the 15 possible pairs per scene, evaluators indicated "Left preferable", "Right preferable", or "Both equal".

Each pair of renderings for the same scene was evaluated by 25 human annotators who indicated which version they preferred or if they considered them equally appealing.

style a	style b	style c	style d	style e	neutral

Rendering Styles

Each scene features six distinct interpretations:

• Expert A–E: Five professional photographer renderings from the original MIT-Adobe FiveK dataset.
• Neutral: Adobe Photoshop’s "Auto" mode rendering, serving as a baseline.

Key Features

• Subjective Focus: Targets the "aesthetics of rendering" (color, texture, and artistic expression) rather than simple signal-to-noise ratios.

• Content-Dependent Preferences: The dataset reveals that preferred rendering styles vary significantly based on image content (e.g., portraits vs. landscapes).

• Personalization Support: Includes unique evaluator identifiers, enabling research into personalized aesthetic preference modeling—a critical area for recommendation systems and generative AI[cite: 1622].

• Content Annotations: Includes scene classifications generated via BLIP (e.g., nature, food, night scenes) with verified 96% accuracy.

Dataset Statistics

Feature	Value
Scenes	5,000
Images	30,000
Annotators	13,648
Votes	2,500,000+
Test Set	1,283 scenes (approx. 25% of data)
Upper Bound Accuracy	0.896 (human consensus)

Applications

REPID is designed to foster research in:

1. Aesthetic Preference Prediction: Training models to predict which of two renderings a human will prefer.
2. Personalized Rendering: Modeling individual user tastes using the provided evaluator IDs.
3. Render Ranking: Developing systems that can automatically select the "best" rendering for a given image.
4. Benchmarking IRQA: Providing a qualitatively different challenge than traditional distortion-based IQA benchmarks.

Citation

@misc{plohotnyuk2025dear,
  title={Beyond distortions — a benchmark for subjective evaluation of image rendering quality},
  author={Plohotnyuk, Vsevolod and Panshin, Artyom and Bani{\'c}, Nikola and Bianco, Simone and Freeman, Michael and Ershov, Egor},
  year={2025},
  eprint={2512.05209},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  note = {Preprint at \url{https://arxiv.org/abs/2512.05209}}
}

License

This project involves two types of content, each subject to different licensing terms:

1. Photos (MIT-Adobe FiveK Dataset)

The original images are provided by MIT and Adobe. You may use these photos for research purposes only.

• Adobe License: The LicenseAdobe.txt applies to all files listed in filesAdobe.txt.
• MIT License: The LicenseAdobeMIT.txt applies to all files listed in filesAdobeMIT.txt.

2. Annotations (My Contributions)

Custom annotations/labels included in this repository are licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Under this license, you are free to:

• Share — copy and redistribute the material in any medium or format.
• Adapt — remix, transform, and build upon the material for any purpose, even commercially.

Under the following terms:

• Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made.

Note: While the annotations themselves allow commercial adaptation, the underlying images from the MIT-Adobe FiveK dataset may still be restricted to non-commercial research use only. Please adhere to the original image licenses accordingly.