Model Card: Pixiv HiRes Performance Testing Dataset

Dataset Name: pixiv_hires_performance_testing

Dataset Version: 1.0

Dataset Summary

The Pixiv HiRes Performance Testing Dataset is a small, high-quality dataset of 1024x1024 resolution images curated from the largest single file size images on Pixiv between 2020 and 2023. These images have been collected and resized to a minimum resolution of 1024x1024 pixels to create a dataset suitable for performance testing and benchmarking of image-based machine learning models, specifically targeting tasks like image synthesis, style transfer, and image-to-image translation.

Dataset Size

Number of Images: 2,000
Total File Size: 2.5 GB

Dataset Composition

Images are sourced from the largest single file size images on Pixiv between 2020 and 2023.
Images have a minimum resolution of 1024x1024 pixels.
The dataset is diverse, containing a variety of image styles and subjects, including illustrations, paintings, and photographs.
All images have been resized to fit the 1024x1024 resolution requirement.

Dataset Preprocessing and Cleaning

Resizing: Images were resized using bicubic interpolation to maintain a consistent resolution of 1024x1024 pixels across the dataset.
Image format: All images have been converted to the lossless PNG format to minimize compression artifacts and preserve image quality.

Dataset Split

This dataset is not split into train, validation, and test sets, as it is intended for performance testing and benchmarking purposes rather than training models. Users are encouraged to create their own splits as needed.

Intended Use

The Pixiv HiRes Performance Testing Dataset is designed for evaluating the performance of image-based machine learning models, particularly those focused on high-resolution image synthesis, style transfer, and image-to-image translation. It can be used for:

Benchmarking and comparing the performance of different models.
Assessing the robustness and generalization capabilities of models when dealing with high-resolution images.
Fine-tuning and optimizing model hyperparameters for high-resolution image tasks.

Limitations and Considerations

The dataset is relatively small and may not cover all possible variations in image content and style.
Images are sourced from Pixiv, which may have a bias towards certain types of art styles and subject matter.
The dataset may contain some images with mature or explicit content. Users should be aware of this when using the dataset for research or applications.

Data Source

Images in this dataset are sourced from Pixiv, specifically the largest single file size images uploaded between 2020 and 2023.