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 (2025), LiblibAI (2025), Civitai (2025)
• Real image source: Unsplash (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:

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} }