Stylebreeder

Dataset Summary

Stylebreeder is a large-scale text-to-image prompt dataset. It contains 6.8 million images generated by various T2I models such as Stable Diffusion, SD-XL, and ControlNet. Images are generated by actual users from Artbreeder.

Stylebreeder is publicly available at 🤗 Hugging Face Dataset.

Supported Tasks and Leaderboards

The massive variety and magnitude of this AI-generated dataset presents opportunities for novel research in generative AI and computer vision.

Two Subsets

Stylebreeder provides two subsets (Stylebreeder 2M and Stylebreeder 6M) to support different needs.

Key Differences

1. Stylebreeder 6M contains many more images and is a superset of Stylebreeder 2M.
2. Images in Stylebreeder 2M are that of the lowest NSFW scores.

Dataset Structure

# Stylebreeder 2M
./
├── 0000.parquet
├── 0001.parquet
├── 0002.parquet
├── 0003.parquet
└── [...]

# Stylebreeder 6M
./
├── 0000.parquet
├── 0001.parquet
├── 0002.parquet
├── 0003.parquet
└── [...]

Data Instances

For example, below is the image of 8ecf2b5411b475b4af5cc728d295.jpeg and its key-value pair in 0019.parquet.

{
  "f3501e05-aef7-4225-a9e9-f516527408ac.png": {
    "prompt": "a landscape painting of alpine golden buckwheat and cats ear plants and valerians in a mangrove swamp, al fresco art by albert goodwin and johan messely and sanford robinson gifford and michal karcz chiaroscuro atmospheric phenomenon, crisp outlines, intricate detail 2",
    "creator_id": 4329037,
    "created_at": 2023-02-19 00:30:24.464962+00:00,
    "modelName": stable-1.5,
    "toxicity": 0.00059,
    "severe_toxicity": 0,
    "obscene": 0.00008,
    "identity_attack": 0.0001,
    "insult": 0.00015,
    "threat": 0.00002,
    "sexual_explicit": 0.00002,
    "image_nsfw": 0,
    "cluster_id": 1245
  },
}

Dataset Metadata

Below are two rows from the metadata.

image_id	image_key	prompt	creator_id	created_at	modelName	toxicity	severe_toxicity	obscene	identity_attack	insult	threat	sexual_explicit	image_nsfw	cluster_id	image
295543512	699f615fa6ce37ac7aab08a629d5	tinytipe child type steve busemi	39096	2024-01-16 01:04:45.261243+00:00	sdlx-1.0	0.00357	0.00001	0.00026	0.00025	0.00041	0.00034	0.00008	0	3164	[IMG]
276545841	f415c6042aba5f32e654de1b71fd	wanted poster, dynamic lighting, tintype daguerrotype, solarized photo print, film grain, pastel color palette, photo by alfred stieglitz and georgia okeeffe	39096	2023-04-21 05:56:45.570525+00:00	stable-1.5	0.0004	0	0.00003	0.00009	0.0001	0.00002	0.00001	0	945	[IMG]

Metadata Schema

Column	Type	Description
image_id	uint16	Image numerical ID.
image_key	string	Image UUID filename.
prompt	string	Text prompt used to generate this image.
creator_id	uint16	ID of user who generated this image.
modelName	string	Model used to generate this image.
toxicity	float32	Toxicity score.
severe_toxicity	float32	Severe toxicity score.
obscene	float32	Obscene score.
identity_attack	float32	Identity attack score.
insult	float32	Insult score.
threat	float32	Threat score.
sexual_explicit	float32	Sexual explicit score.
image_nsfw	float32	NSFW score.
cluster_id	float32	Assigned style cluster ID.
artists	string	Artists in the generated image.
image	float32	Image file.

Method 1: Using Hugging Face Datasets Loader

The Hugging Face Datasets library provides functions to easily load prompts and images from Stylebreeder. You can see all subsets in the Dataset Preview.

import numpy as np
from datasets import load_dataset
# Load the dataset with the `2M_sample` subset
dataset = load_dataset('stylebreeder/stylebreeder', '2M_sample')

Method 2. Use the scraping script

import os 
from datasets import load_dataset

image_folder = "./all_images"
os.makedirs(image_folder, exists_ok=True)

def download_image(image_key):
    image_url = f"https://artbreeder.b-cdn.net/imgs/{image_key}.jpeg"
    save_path = os.path.join(image_folder, f"{image_key}.jpeg")

    response = requests.get(image_url)
    if response.status_code == 200:
        with open(save_path, 'wb') as file:
            file.write(response.content)
    else:
        print(f"Failed to download image: {image_key}")


dataset = load_dataset("stylebreeder/stylebreeder", split='6M_full')
dd = dataset.select_columns(['image_key']).to_pandas()
dd.image_key.apply(lambda x: download_image(x))

Dataset Creation

Curation Rationale

As artists worldwide are increasingly leveraging text-to-image diffusion models to create artworks spanning diverse set of styles, there is a pressing need uncover and promote unique artistic expressions. This will enable users to discover unique artistic styles, generate personalized content using those styles, and get recommendations about styles they would be interested. Stylebreeder was created to address these questions, and to the best of our knowledge, has the largest userbase (95K) spanning to 6.8M images with 1.8M unique text prompts.

Collection Process

The images, prompt and other metadata including including Positive Prompt, Negative Prompt, anonymized ImageID, anonymized UserID, Timestamp, and Image Size (height, width), and model-related hyperparameters including Model Type, Seed, Step, and CFG Scale are collected from Artbreeder website. Moreover we derived additional features such as Cluster ID, Prompt NSFW, Image NSFW, Toxicity, Insult, Threat, Identity Attack scores, Artist names.

Data Instances

Each instance consists of an image generated by a text-to-image diffusion model (such as SD-XL or ControlNet) and the original text prompt used for generating the image, as well as original metadata and expanded features created by us. The metadata originally collected from the Artbreeder website are: Positive Prompt, Negative Prompt, anonymized ImageID, anonymized UserID, Timestamp, and Image Size (height, width), and model-related hyperparameters including Model Type, Seed, Step, and CFG Scale. We also offer further metadata like Cluster ID, along with scores for Prompt NSFW, Image NSFW, Toxicity, Insult, Threat, Identity Attack. We also provide potential artist names that are used in the text prompt such as “Ilya Kuvshinov”.

Who are the source producers?

The original producers of images are users of Artbreeder.

Annotations

No additional annotations.

Personal and Sensitive Information

We share anonymized user IDs and text prompts may have celebrity or artist names in the text prompts.

Considerations for Using the Data

Biases

From a societal perspective, while our tools aim to democratize art creation, there exists a risk of reinforcing existing biases present in the training data, which could skew the diversity and representation in generated artworks.

Other Limitations

One of the primary limitations lies in the potential for over-reliance on technology in artistic creation, which could diminish the value and perception of human-driven artistry and creativity. Additionally, the use of AI in art generation raises concerns about copyright and originality, especially when styles closely mimic those of existing artists without clear attribution.

Additional Information

We provide a Google form for reporting harmful or inappropriate images and prompts, as well as allowing artists to opt-out in case their names are used in the text prompts.

Licensing Information

The Stylebreeder dataset is available under the CC0 1.0 License.