Table of Contents

Dataset Description

Dataset Summary

DiffusionDB is the first large-scale text-to-image prompt dataset. It contains 14 million images generated by Stable Diffusion using prompts and hyperparameters specified by real users.

DiffusionDB is publicly available at 🤗 Hugging Face Dataset.

Supported Tasks and Leaderboards

The unprecedented scale and diversity of this human-actuated dataset provide exciting research opportunities in understanding the interplay between prompts and generative models, detecting deepfakes, and designing human-AI interaction tools to help users more easily use these models.


The text in the dataset is mostly English. It also contains other languages such as Spanish, Chinese, and Russian.

Two Subsets

DiffusionDB provides two subsets (DiffusionDB 2M and DiffusionDB Large) to support different needs.

Subset Num of Images Num of Unique Prompts Size Image Directory Metadata Table
DiffusionDB 2M 2M 1.5M 1.6TB images/ metadata.parquet
DiffusionDB Large 14M 1.8M 6.5TB diffusiondb-large-part-1/ diffusiondb-large-part-2/ metadata-large.parquet
Key Differences
  1. Two subsets have a similar number of unique prompts, but DiffusionDB Large has much more images. DiffusionDB Large is a superset of DiffusionDB 2M.
  2. Images in DiffusionDB 2M are stored in png format; images in DiffusionDB Large use a lossless webp format.

Dataset Structure

We use a modularized file structure to distribute DiffusionDB. The 2 million images in DiffusionDB 2M are split into 2,000 folders, where each folder contains 1,000 images and a JSON file that links these 1,000 images to their prompts and hyperparameters. Similarly, the 14 million images in DiffusionDB Large are split into 14,000 folders.

# DiffusionDB 2M
├── images
│   ├── part-000001
│   │   ├── 3bfcd9cf-26ea-4303-bbe1-b095853f5360.png
│   │   ├── 5f47c66c-51d4-4f2c-a872-a68518f44adb.png
│   │   ├── 66b428b9-55dc-4907-b116-55aaa887de30.png
│   │   ├── [...]
│   │   └── part-000001.json
│   ├── part-000002
│   ├── part-000003
│   ├── [...]
│   └── part-002000
└── metadata.parquet
# DiffusionDB Large
├── diffusiondb-large-part-1
│   ├── part-000001
│   │   ├── 0a8dc864-1616-4961-ac18-3fcdf76d3b08.webp
│   │   ├── 0a25cacb-5d91-4f27-b18a-bd423762f811.webp
│   │   ├── 0a52d584-4211-43a0-99ef-f5640ee2fc8c.webp
│   │   ├── [...]
│   │   └── part-000001.json
│   ├── part-000002
│   ├── part-000003
│   ├── [...]
│   └── part-010000
├── diffusiondb-large-part-2
│   ├── part-010001
│   │   ├── 0a68f671-3776-424c-91b6-c09a0dd6fc2d.webp
│   │   ├── 0a0756e9-1249-4fe2-a21a-12c43656c7a3.webp
│   │   ├── 0aa48f3d-f2d9-40a8-a800-c2c651ebba06.webp
│   │   ├── [...]
│   │   └── part-000001.json
│   ├── part-010002
│   ├── part-010003
│   ├── [...]
│   └── part-014000
└── metadata-large.parquet

These sub-folders have names part-0xxxxx, and each image has a unique name generated by UUID Version 4. The JSON file in a sub-folder has the same name as the sub-folder. Each image is a PNG file (DiffusionDB 2M) or a lossless WebP file (DiffusionDB Large). The JSON file contains key-value pairs mapping image filenames to their prompts and hyperparameters.

Data Instances

For example, below is the image of f3501e05-aef7-4225-a9e9-f516527408ac.png and its key-value pair in part-000001.json.

  "f3501e05-aef7-4225-a9e9-f516527408ac.png": {
    "p": "geodesic landscape, john chamberlain, christopher balaskas, tadao ando, 4 k, ",
    "se": 38753269,
    "c": 12.0,
    "st": 50,
    "sa": "k_lms"

Data Fields

  • key: Unique image name
  • p: Prompt
  • se: Random seed
  • c: CFG Scale (guidance scale)
  • st: Steps
  • sa: Sampler

Dataset Metadata

To help you easily access prompts and other attributes of images without downloading all the Zip files, we include two metadata tables metadata.parquet and metadata-large.parquet for DiffusionDB 2M and DiffusionDB Large, respectively.

The shape of metadata.parquet is (2000000, 13) and the shape of metatable-large.parquet is (14000000, 13). Two tables share the same schema, and each row represents an image. We store these tables in the Parquet format because Parquet is column-based: you can efficiently query individual columns (e.g., prompts) without reading the entire table.

Below are three random rows from metadata.parquet.

image_name prompt part_id seed step cfg sampler width height user_name timestamp image_nsfw prompt_nsfw
0c46f719-1679-4c64-9ba9-f181e0eae811.png a small liquid sculpture, corvette, viscous, reflective, digital art 1050 2026845913 50 7 8 512 512 c2f288a2ba9df65c38386ffaaf7749106fed29311835b63d578405db9dbcafdb 2022-08-11 09:05:00+00:00 0.0845108 0.00383462
a00bdeaa-14eb-4f6c-a303-97732177eae9.png human sculpture of lanky tall alien on a romantic date at italian restaurant with smiling woman, nice restaurant, photography, bokeh 905 1183522603 50 10 8 512 768 df778e253e6d32168eb22279a9776b3cde107cc82da05517dd6d114724918651 2022-08-19 17:55:00+00:00 0.692934 0.109437
6e5024ce-65ed-47f3-b296-edb2813e3c5b.png portrait of barbaric spanish conquistador, symmetrical, by yoichi hatakenaka, studio ghibli and dan mumford 286 1713292358 50 7 8 512 640 1c2e93cfb1430adbd956be9c690705fe295cbee7d9ac12de1953ce5e76d89906 2022-08-12 03:26:00+00:00 0.0773138 0.0249675

Metadata Schema

metadata.parquet and metatable-large.parquet share the same schema.

Column Type Description
image_name string Image UUID filename.
prompt string The text prompt used to generate this image.
part_id uint16 Folder ID of this image.
seed uint32 Random seed used to generate this image.
step uint16 Step count (hyperparameter).
cfg float32 Guidance scale (hyperparameter).
sampler uint8 Sampler method (hyperparameter). Mapping: {1: "ddim", 2: "plms", 3: "k_euler", 4: "k_euler_ancestral", 5: "k_heun", 6: "k_dpm_2", 7: "k_dpm_2_ancestral", 8: "k_lms", 9: "others"}.
width uint16 Image width.
height uint16 Image height.
user_name string The unique discord ID’s SHA256 hash of the user who generated this image. For example, the hash for xiaohk#3146 is e285b7ef63be99e9107cecd79b280bde602f17e0ca8363cb7a0889b67f0b5ed0. “deleted_account” refer to users who have deleted their accounts. None means the image has been deleted before we scrape it for the second time.
timestamp timestamp UTC Timestamp when this image was generated. None means the image has been deleted before we scrape it for the second time. Note that timestamp is not accurate for duplicate images that have the same prompt, hypareparameters, width, height.
image_nsfw float32 Likelihood of an image being NSFW. Scores are predicted by LAION’s state-of-art NSFW detector (range from 0 to 1). A score of 2.0 means the image has already been flagged as NSFW and blurred by Stable Diffusion.
prompt_nsfw float32 Likelihood of a prompt being NSFW. Scores are predicted by the library Detoxicy. Each score represents the maximum of toxicity and sexual_explicit (range from 0 to 1).

Warning Although the Stable Diffusion model has an NSFW filter that automatically blurs user-generated NSFW images, this NSFW filter is not perfect—DiffusionDB still contains some NSFW images. Therefore, we compute and provide the NSFW scores for images and prompts using the state-of-the-art models. The distribution of these scores is shown below. Please decide an appropriate NSFW score threshold to filter out NSFW images before using DiffusionDB in your projects.

Data Splits

For DiffusionDB 2M, we split 2 million images into 2,000 folders where each folder contains 1,000 images and a JSON file. For DiffusionDB Large, we split 14 million images into 14,000 folders where each folder contains 1,000 images and a JSON file.

Loading Data Subsets

DiffusionDB is large (1.6TB or 6.5 TB)! However, with our modularized file structure, you can easily load a desirable number of images and their prompts and hyperparameters. In the example-loading.ipynb notebook, we demonstrate three methods to load a subset of DiffusionDB. Below is a short summary.

Method 1: Using Hugging Face Datasets Loader

You can use the Hugging Face Datasets library to easily load prompts and images from DiffusionDB. We pre-defined 16 DiffusionDB subsets (configurations) based on the number of instances. You can see all subsets in the Dataset Preview.

import numpy as np
from datasets import load_dataset

# Load the dataset with the `large_random_1k` subset
dataset = load_dataset('poloclub/diffusiondb', 'large_random_1k')

Method 2. Use the PoloClub Downloader

This repo includes a Python downloader download.py that allows you to download and load DiffusionDB. You can use it from your command line. Below is an example of loading a subset of DiffusionDB.


The script is run using command-line arguments as follows:

  • -i --index - File to download or lower bound of a range of files if -r is also set.
  • -r --range - Upper bound of range of files to download if -i is set.
  • -o --output - Name of custom output directory. Defaults to the current directory if not set.
  • -z --unzip - Unzip the file/files after downloading
  • -l --large - Download from Diffusion DB Large. Defaults to Diffusion DB 2M.
Downloading a single file

The specific file to download is supplied as the number at the end of the file on HuggingFace. The script will automatically pad the number out and generate the URL.

python download.py -i 23
Downloading a range of files

The upper and lower bounds of the set of files to download are set by the -i and -r flags respectively.

python download.py -i 1 -r 2000

Note that this range will download the entire dataset. The script will ask you to confirm that you have 1.7Tb free at the download destination.

Downloading to a specific directory

The script will default to the location of the dataset’s part .zip files at images/. If you wish to move the download location, you should move these files as well or use a symbolic link.

python download.py -i 1 -r 2000 -o /home/$USER/datahoarding/etc

Again, the script will automatically add the / between the directory and the file when it downloads.

Setting the files to unzip once they’ve been downloaded

The script is set to unzip the files after all files have downloaded as both can be lengthy processes in certain circumstances.

python download.py -i 1 -r 2000 -z

Method 3. Use metadata.parquet (Text Only)

If your task does not require images, then you can easily access all 2 million prompts and hyperparameters in the metadata.parquet table.

from urllib.request import urlretrieve
import pandas as pd

# Download the parquet table
table_url = f'https://huggingface.co/datasets/poloclub/diffusiondb/resolve/main/metadata.parquet'
urlretrieve(table_url, 'metadata.parquet')

# Read the table using Pandas
metadata_df = pd.read_parquet('metadata.parquet')

Dataset Creation

Curation Rationale

Recent diffusion models have gained immense popularity by enabling high-quality and controllable image generation based on text prompts written in natural language. Since the release of these models, people from different domains have quickly applied them to create award-winning artworks, synthetic radiology images, and even hyper-realistic videos.

However, generating images with desired details is difficult, as it requires users to write proper prompts specifying the exact expected results. Developing such prompts requires trial and error, and can often feel random and unprincipled. Simon Willison analogizes writing prompts to wizards learning “magical spells”: users do not understand why some prompts work, but they will add these prompts to their “spell book.” For example, to generate highly-detailed images, it has become a common practice to add special keywords such as “trending on artstation” and “unreal engine” in the prompt.

Prompt engineering has become a field of study in the context of text-to-text generation, where researchers systematically investigate how to construct prompts to effectively solve different down-stream tasks. As large text-to-image models are relatively new, there is a pressing need to understand how these models react to prompts, how to write effective prompts, and how to design tools to help users generate images. To help researchers tackle these critical challenges, we create DiffusionDB, the first large-scale prompt dataset with 14 million real prompt-image pairs.

Source Data

Initial Data Collection and Normalization

We construct DiffusionDB by scraping user-generated images on the official Stable Diffusion Discord server. We choose Stable Diffusion because it is currently the only open-source large text-to-image generative model, and all generated images have a CC0 1.0 Universal Public Domain Dedication license that waives all copyright and allows uses for any purpose. We choose the official Stable Diffusion Discord server because it is public, and it has strict rules against generating and sharing illegal, hateful, or NSFW (not suitable for work, such as sexual and violent content) images. The server also disallows users to write or share prompts with personal information.

Who are the source language producers?

The language producers are users of the official Stable Diffusion Discord server.


The dataset does not contain any additional annotations.

Annotation process


Who are the annotators?


Personal and Sensitive Information

The authors removed the discord usernames from the dataset. We decide to anonymize the dataset because some prompts might include sensitive information: explicitly linking them to their creators can cause harm to creators.

Considerations for Using the Data

Social Impact of Dataset

The purpose of this dataset is to help develop better understanding of large text-to-image generative models. The unprecedented scale and diversity of this human-actuated dataset provide exciting research opportunities in understanding the interplay between prompts and generative models, detecting deepfakes, and designing human-AI interaction tools to help users more easily use these models.

It should note that we collect images and their prompts from the Stable Diffusion Discord server. The Discord server has rules against users generating or sharing harmful or NSFW (not suitable for work, such as sexual and violent content) images. The Stable Diffusion model used in the server also has an NSFW filter that blurs the generated images if it detects NSFW content. However, it is still possible that some users had generated harmful images that were not detected by the NSFW filter or removed by the server moderators. Therefore, DiffusionDB can potentially contain these images. To mitigate the potential harm, we provide a Google Form on the DiffusionDB website where users can report harmful or inappropriate images and prompts. We will closely monitor this form and remove reported images and prompts from DiffusionDB.

Discussion of Biases

The 14 million images in DiffusionDB have diverse styles and categories. However, Discord can be a biased data source. Our images come from channels where early users could use a bot to use Stable Diffusion before release. As these users had started using Stable Diffusion before the model was public, we hypothesize that they are AI art enthusiasts and are likely to have experience with other text-to-image generative models. Therefore, the prompting style in DiffusionDB might not represent novice users. Similarly, the prompts in DiffusionDB might not generalize to domains that require specific knowledge, such as medical images.

Other Known Limitations

Generalizability. Previous research has shown a prompt that works well on one generative model might not give the optimal result when used in other models. Therefore, different models can need users to write different prompts. For example, many Stable Diffusion prompts use commas to separate keywords, while this pattern is less seen in prompts for DALL-E 2 or Midjourney. Thus, we caution researchers that some research findings from DiffusionDB might not be generalizable to other text-to-image generative models.

Additional Information

Dataset Curators

DiffusionDB is created by Jay Wang, Evan Montoya, David Munechika, Alex Yang, Ben Hoover, Polo Chau.

Licensing Information

The DiffusionDB dataset is available under the CC0 1.0 License. The Python code in this repository is available under the MIT License.

Citation Information

  title = :  Large-Scale Prompt Gallery Dataset for Text-to-Image Generative Models},
  author = {Wang, Zijie J. and Montoya, Evan and Munechika, David and Yang, Haoyang and Hoover, Benjamin and Chau, Duen Horng},
  year = {2022},
  journal = {arXiv:2210.14896 [cs]},
  url = {https://arxiv.org/abs/2210.14896}


If you have any questions, feel free to open an issue or contact Jay Wang.