| --- |
| license: cc-by-nc-4.0 |
| --- |
| # HiRes-50K Dataset |
|
|
| **HiRes-50K** is a **cross-domain evaluation-only dataset** designed to assess the **generalization capability of AI-generated image (AIGI) detection models** and their performance on **high-resolution, high-fidelity images**. This dataset is not intended for model training and should only be used for evaluation purposes. |
|
|
|
|
| ## Dataset Overview |
|
|
| HiRes-50K consists of **50,568 images**, covering long-edge resolutions from below 1K to over 10K pixels, with some reaching up to 64 megapixels. |
|
|
| The dataset is collected from the following publicly accessible communities: |
|
|
| - **AI-generated image sources**: [Freepik](https://www.freepik.com/) (2025), [LiblibAI](https://www.liblib.art/) (2025), [Civitai](https://civitai.com/) (2025) |
| - **Real image source**: [Unsplash](https://unsplash.com/) (2025) |
|
|
| All images were collected in compliance with the Terms of Service and Privacy Policies of their respective sources at the time of access. |
|
|
| ## Dataset Composition |
|
|
| ### 1. AI-Generated Images |
|
|
| - **Quantity**: ~25,000 images |
| - **Content categories**: |
| - Portraits (close-ups, upper-body, full-body, and group images) |
| - Landscapes (mountains, beaches, cities, rural areas, deserts, various weather conditions) |
| - Architecture (urban scenes, skyscrapers, villas, neighborhoods) |
| - Vehicles and animals |
|
|
| **Resolution distribution:** |
|
|
| | Resolution range (px, long edge) | Image count | |
| | -------------------------------- | ----------- | |
| | [0, 900) | 845 | |
| | [900, 1200) | 6,665 | |
| | [1200, 1500) | 6,399 | |
| | [1500, 2000) | 5,262 | |
| | [2000, 2500) | 3,674 | |
| | [2500, 3000) | 571 | |
| | [3000, 5000) | 1,196 | |
| | [5000, ∞) | 472 | |
|
|
| All images were filtered to ensure high JPEG quality (quality factor ≥ 75). |
|
|
| ### 2. Real Images |
|
|
| To ensure a fair comparison, real images were matched with AI-generated images in both **resolution** and **JPEG compression level**. Real images were resized to match the pixel count of their synthetic counterparts while preserving aspect ratios. JPEG compression was applied with identical quality settings |
|
|
| ## Citation |
|
|
| If you use this dataset in your research, please cite the following paper: |
|
|
| > @article{zhang2025nopixel, |
| > title={No Pixel Left Behind: A Detail-Preserving Architecture for Robust High-Resolution AI-Generated Image Detection}, |
| > author={Lianrui Mu, Zou Xingze, Jianhong Bai, and others}, |
| > journal={arXiv preprint arXiv:2508.17346}, |
| > year={2025}, |
| > url={https://arxiv.org/abs/2508.17346} |
| > } |
