Dataset Card for Midjourney-Threads 🧵💬

This dataset contains users prompts from the Midjourney discord channel, organized into "threads of interaction". Each thread contains a user’s trails to create one target image.

The dataset was introduced as part of the paper: Human Learning by Model Feedback: The Dynamics of Iterative Prompting with Midjourney.

Dataset Sources

• Repository: https://github.com/shachardon/Mid-Journey-to-alignment
• Paper: https://aclanthology.org/2023.emnlp-main.253/

Dataset Structure

Main Columns:

• 'text' - the original prompt
• 'args' - predefined parameters (such as the aspect ratio, chaos and more)
• 'channel_id' - the discord channel
• 'userid' - an anonymous user id
• 'timestamp' - a timestamp of the prompt creation
• 'label' - Ture whether an image that was generated based on that prompt was upscaled, otherwise False.
• 'id' - unique id of the prompt
• 'url_png' - link to the generated images (a 4-grid version)
• 'main_content' - prefix of the prompt, without trailing magic-words
• 'concreteness' - concreteness score, based on the this paper
• 'word_len' - the number of words
• 'repeat_words' - the occurrences of each word that appears more than once in the prompt, excluding stop words.
• 'reapeat_words_ratio' - repeat_words / word_len
• 'perplexity' - the perplexity GPT-2 assigns to each prompt.
• 'caption_0-3' - captions that were generated by the BLIP-2 model, with the 4 created images as its inputs.
• 'phase' - train/test split, as was used to train image/text classifiers
• 'magic_ratio' - the percentage of words that were recognized as magic words in the prompt
• 'thread_id' - the id of the thread
• 'depth' - the max depth of a constituency parse tree of the prompt.
• 'num_sent_parser' - the number of sentences in the prompt.
• 'num_sent_parser_ratio' - num_sent_parser / word_len
• 'words_per_sent' - word_len / num_sent_parser

Dataset Creation

Source Data

We construct the dataset by scraping user-generated prompts from the Midjourney Discord server. The server contains channels in which a user can type a prompt and arguments, and then the Midjourney bot would reply with 4 generated images, combined together into a grid. Then, if the user is satisfied with one of the 4 images, they can send an 'upscale' command to the bot, to get an upscaled version of the desired image. We randomly choose one of the 'newbies' channels, where both new and experienced users are experimenting with general domain prompts. We collect 693,528 prompts (From 23 January to 1 March 2023), together with their matching images and meta-data such as timestamps and user ids (which we anonymize).

Data Collection and Processing

We split the prompts into threads automatically, see the paper for more details. In addition, we extract features (perplexity, sentence length, and more).

Personal and Sensitive Information

We fully anonymize the data by removing user names and other user-specific meta-data. If you recognize your prompts here and want to remove them, please send us an email.

The Midjourney Discord is an open community that allows others to use images and prompts whenever they are posted in a public setting. Paying users do own all assets they create, and therefore we do not include the image files in our dataset, but only links to them.

Recommendations, Risks, and Limitations

We split the prompts into threads automatically, and therefore there are some mistakes. For more about our annotations method, please see the paper. Our manual sample did not find any offensive content in the prompts.

Citation

BibTeX:

@inproceedings{don-yehiya-etal-2023-human,
    title = "Human Learning by Model Feedback: The Dynamics of Iterative Prompting with Midjourney",
    author = "Don-Yehiya, Shachar  and
      Choshen, Leshem  and
      Abend, Omri",
    editor = "Bouamor, Houda  and
      Pino, Juan  and
      Bali, Kalika",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.emnlp-main.253",
    pages = "4146--4161",
    abstract = "Generating images with a Text-to-Image model often requires multiple trials, where human users iteratively update their prompt based on feedback, namely the output image. Taking inspiration from cognitive work on reference games and dialogue alignment, this paper analyzes the dynamics of the user prompts along such iterations. We compile a dataset of iterative interactions of human users with Midjourney. Our analysis then reveals that prompts predictably converge toward specific traits along these iterations. We further study whether this convergence is due to human users, realizing they missed important details, or due to adaptation to the model{'}s {``}preferences{''}, producing better images for a specific language style. We show initial evidence that both possibilities are at play. The possibility that users adapt to the model{'}s preference raises concerns about reusing user data for further training. The prompts may be biased towards the preferences of a specific model, rather than align with human intentions and natural manner of expression.",
}