AnimePics

This dataset is a pure image dataset in WebDataset format, designed for pre-training anime-style models. It can also be used to evaluate the effect of additional pre-training on backbones trained primarily on real-world images when applied to datasets of a completely different nature. The dataset is designed to be continuously updated, leveraging the features of WebDataset.

Dataset Details

Dataset Description

The animepics dataset is a large-scale, image-only dataset intended for unsupervised pre-training of models for anime-style image generation and analysis. All images are provided in the WebDataset format, allowing for efficient, sequential access to the data, which is particularly well-suited for large-scale training on cloud infrastructure.

• Curated by: zenless-lab
• License: MIT

Dataset Sources [optional]

• Repository: zenless-lab/animepics

Uses

Direct Use

This dataset is intended for pre-training computer vision models, particularly those that will be used for tasks involving anime-style images. It can also be used to research the effects of transfer learning from a distinctly different domain (anime) to models pre-trained on real-world images.

Out-of-Scope Use

This dataset is not intended for any use that would violate the original licenses of the source datasets. The images are provided AS-IS, and may contain biases or content that is not suitable for all audiences.

Dataset Structure

The dataset is structured in the WebDataset format. Each .tar file contains a collection of .webp images, with a maximum size of 1GB per .tar file. The key for each image (which also serves as its filename) is a UUID4(e.g. bf8f5936-6b93-489f-9e07-01bbb7df8db7), ensuring that there are no filename conflicts when new images are added. This structure allows for the dataset to be continuously and sustainably updated.

Dataset Creation

Curation Rationale

The motivation behind creating this dataset was to provide a large-scale, easily accessible, and continuously updated resource for pre-training models on anime-style images.

Source Data

Data Collection and Processing

The images in this dataset are sourced from other open datasets with permissive licenses that are compatible with the MIT license, allowing for redistribution.

The current data source is:

• nyanko7/danbooru2023

All images have been pre-processed for efficient training on cloud infrastructure. Images with a short side larger than 1024 pixels have been resized to have a short side of 1024 pixels while maintaining their original aspect ratio. All images have been converted to the .webp format.

The dataset currently contains 5,954,808 images, with a total size of 640GB.