license: cc0-1.0
task_categories:
- text-to-image
language:
- en
Microcosmos Dataset
This dataset consists of a carefully curated collection of Creative Commons (CC0) images or similar, combined with both synthetic and human-generated captions. It was assembled to facilitate the training of diffusion models with a focus on efficiency and ethical data practices. The dataset was compiled over several months, highlighting the dedication to responsible data collection and management. Dataset Details Dataset Description Microcosmos is designed to train high-performance diffusion models using a smaller, yet diverse, dataset. It includes high-resolution 768x768 images sourced from various Creative Commons platforms. These images are complemented by a mix of synthetic and human-generated captions, providing a robust foundation for generating and evaluating images.
Data Sources and License
The images in this dataset have been sourced from the following platforms:
Pexels: pexels.com Unsplash: unsplash.com opengameart.org: opengameart.org Met Museum Open Access: metmuseum.org KissCC0: kisscc0.com LIBRESHOT libreshot.com
Although the dataset primarily consists of CC0 and permissive content, only images that explicitly permit redistribution will be made available. In alignment with site policies and creators' preferences, images without explicit redistribution permissions will be indexed but not redistributed. If you are the rights holder and believe an image has been included in error or wish to raise concerns, please reach out to us via email for review. Non-redistributable content: https://unsplash.com/data/lite/latest
Many old car models no longer have active patents, allowing their use in creations, such as generative AI models, without infringing intellectual property rights. No explicit images of cars or aircraft have been added to the training so far, but if they are, they will be considered car models until the 2000s. In Brazil, the protection of a vehicle design is granted through an industrial design registration, and its duration is different from invention patents. According to the Industrial Property Law (Law No. 9,279/1996), industrial designs are protected for an initial period of 10 years from the date of filing. This period can be extended for up to three consecutive periods of 5 years each, totaling a maximum of 25 years of protection.
AI-generated images present legal uncertainties in Brazil, particularly concerning the country’s copyright law (LDA) and the General Data Protection Law (LGPD), which restricts the use of personal data in training algorithms. Whether AI-generated images are considered derivative or transformative will determine their legal standing and potential use in training other models. To avoid conflicts, AI-generated content has not been included in this dataset. As AI-generated images become more common many free content sites, such as those used in research so there may be the unintentional appearance of this type of content. If any AI-generated images are included, they must meet the criteria of being transformative, original, and non-individualized, as per LGPD and LDA requirements.
Dataset Structure
The dataset is organized into multiple subsets, each containing images categorized by a theme. The dataset's structure ensures efficient training of diffusion models while maintaining a balance between image diversity and dataset size. This organization helped in the process of mixed captions.
NLP: English (en)