README.md

metadata

license: mit
task_categories:
  - text-to-image
  - image-to-text
tags:
  - art
  - furry
  - e621
  - twitter
  - instagram
  - jtp-3
size_categories:
  - 10K<n<100K
configs:
  - config_name: twitter
    data_files:
      - split: train
        path: xfiles/**/*.jpg
      - split: train
        path: xfiles/**/*.png
  - config_name: instagram
    data_files:
      - split: train
        path: Instagram/**/*.jpg
      - split: train
        path: Instagram/**/*.png

viewer: true

Dataset Card for TwitterArtists (Pixel-Dust)

This dataset is a collection of art and media scraped from various artists and profiles across X (formerly Twitter) and Instagram. It is primarily focused on furry art and similar stylized content, intended for use in training or fine-tuning generative models.

Dataset Description

• Repository: Pixel-Dust/TwitterArtists
• Content Type: Images & Tags
• Primary Focus: Furry artists, digital illustrations, and character art.
• Sources: X (Twitter), Instagram.
• License: MIT

Data Collection & Annotation

Source Data

The images were collected from social media profiles of numerous artists. While the bulk of the content is high-quality digital art, the collection process was broad, it contain some +18 content.

Captioning & Tagging

All images have been automatically captioned using JTP-3 (Joint Tagger Project).

• Tagging Style: the tags follow the e621-style format (shorthand descriptive tags, species, character traits, and composition metadata).
• Format: Typically stored as .txt files accompanying each image.

Known Limitations & "Noise"

Please be aware: This dataset is currently in a "raw" or "semi-filtered" state.

• Non-Art Content: You may find "trash" media, including repeated memes, screenshots, or other non-illustration files that were scraped alongside the art.
• Redundancy: There are known instances of duplicate or near-duplicate images.
• Tag Accuracy: While JTP-3 is robust, automated tagging is not 100% accurate and may hallucinate tags or miss specific character details, can have false positives.

Call for Contributions

I am actively looking for help to clean and refine this dataset. If you are interested in:

1. Filtering: Removing memes, duplicates, or low-quality "noise."
2. Tag Improvement: Refining the e621-style tags or adding missing metadata.
3. Curating: Sorting the content into specific sub-categories or styles.

Attempts to improve the quality of this dataset are highly appreciated! Please feel free to open a Discussion or pull request if you have a filtered version or improved tag set.

Disclaimer

This dataset is provided for research and generative AI training purposes. Please don't respect the original artists' rights and use this data responsibly. Be aware that some of the profiles scraped have sensitive and personal information that the people from X like to expose to everyone. I don't know why they do that, but if you see such information in this dataset, this is not a leak or an invasion of privacy, as all this information is publicly available on X for anyone to see. This dataset does not aim to expose anyone. Please contact me for removal alongside any other trash you find.