The dataset can be downloaded at https://archive.org/download/stack-exchange-images (~799 GiB).

The dataset contains all the ~8.7 million images hosted on i.stack.imgur.com that are linked to in the posts, post histories and comments from the Stack Exchange Data Dump dated 2023-12-08.

The dataset is licensed under Creative Commons Attribution-ShareAlike, following the license used for Stack Exchange user content.

Each filename on i.stack.imgur.com has the following format: 5 chars followed by the file extension.

• Allowed chars: 0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ (10 digits + 26 letters of the alphabet * 2 cases = 62 allowed chars)
• Allowed extension: .gif, .jpg and .png.

The i.stack.imgur.com filenames as well as the Last-Modified response HTTP header were kept. Note that the 5 chars are enough to identify an image, i.e. two images may not share the same 5 chars (e.g., it's not possible to have both hello.jpg and hello.png).

The dataset comprises 62 zip files: the zip file where an image is stored corresponds to the first letter if the filename (0.zip, 1.zip, ..., 9.zip, a.zip, b.zip, ..., .zip, A.zip, B.zip, ..., .Z.zip). Within the zip files, the folder structure is as followed: the first letter of the filename is used for the level-1 folder, and the second letter of the filename is used for the level-2 folder.

Example:

.
├── A
│   ├── 0
│   │   ├── A0gpf.png
│   │   ├── A0mP1.jpg
│   │   ├── A0z99.png
│   ├── 1
│   │   ├── A1Aaq.gif
│   │   ├── A1pdf.png
│   │   ├── A1t51.png
├── M
│   ├── b
│   │   ├───Mb5AA.jpg
│   │   ├── Mbm0z.jpg
│   │   ├── Mbzas.gif
│   ├── Z
│   │   ├───MZ0zc.gif
│   │   ├── MZ546.gif
│   │   ├── MZa12.jpg

Important: one may run into issues using the dataset with Windows because the filenames in this dataset are case-sensitive (following i.stack.imgur.com naming convention), while Windows is mostly case-insensitive. E.g., if one uncompresses a zip file containing the files hello.png and Hello.png, only one file will remain once the extraction is completed. To use the dataset on Windows, one strategy would be to rename files (e.g., adding an underscore after any lowercase letter).

Total number of images: 8,737,759 images. Size: 857,964,647,071 bytes (~799 GiB, i.e. ~858 GB).

Image count by extension (using this command):

• .gif: 206,096 images
• .jpg: 1,675,282 images
• .png: 6,856,303 images
• 78 files have some other extension (8,737,759 - (206,096 +1,675,282 + 6,856,303) = 78). I don't know where they came from, but it's rather negligible so let's just ignore them for now.

Over 99.9% of the files are below 2 MiB, since that's the upload limit on SE, however a few files are above 2 MiB. They can be found with this command.

If you use this dataset in your research, please cite:

@misc{StackExchangeImageDataset,
  author = {Dernoncourt, Franck},
  title = {Stack Exchange Image Dataset},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  doi = {10.5281/zenodo.10465913},
  howpublished = {\url{https://github.com/Franck-Dernoncourt/Stack-Exchange-Image-Dataset}}
}