You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This repository contains an archived subset of X-AIGD real images for reproducibility. Access is restricted to non-commercial academic use.
Log in or Sign Up to review the conditions and access this dataset content.
X-AIGD (real images)
Dataset Structure
The X-AIGD dataset is split into two components:
- Generated Images: Available publicly in the main X-AIGD repo.
- Real Images: The metadata and download script are provided in the main repo. A subset of image copies is hosted in this gated repository for research reproducibility.
Real Image Sources
- COCO: Provided here as an archive for research reproducibility. Note that the image copies here differ from the original ones due to an extra JPEG compression step.
- LAION-2B-en-aesthetic: Provided here as an archive for research reproducibility.
- CC3M: Provided as an archive for research reproducibility.
- SA-1B: Due to strict licensing restrictions by Meta AI, we do not redistribute SA-1B images. Please use the download script provided in the main repo to fetch these images directly from the official source.
Reconstruction of the Full Real-Image Set
To reconstruct the full X-AIGD real-image set exactly, we will first materialize the non-SA-1B images from this gated repository into a local directory, then use the main X-AIGD downloader to fetch the remaining SA-1B images. The final result is a single directory containing all real images, each named as <uid>.<format>.
Step 0: Prepare the environment
Install the required Python packages:
pip install datasets huggingface_hub pillow
Step 1: Export the non-SA-1B images from this gated repo
Choose a local directory that will eventually contain the full real-image set, for example ./real_images, then run:
python export_real_images_from_hf_hub.py \
--output-dir ./real_images
This step writes the images sourcing from MSCOCO, LAION-2B-en-aesthetic, and CC3M into ./real_images using filenames of the form <uid>.<format>.
Step 2: Get the metadata and downloader from the main X-AIGD repo
Download the following files from the main X-AIGD repo:
real_image_metadata.csvdownload_real_images.py
Step 3: Download the remaining SA-1B images
First, download the official index file (text file with a list of all links) from Meta: https://ai.meta.com/datasets/segment-anything-downloads/
Then run the main X-AIGD downloader with the same output directory and pass the local path to that index file:
python download_real_images.py \
--metadata real_image_metadata.csv \
--output-dir ./real_images \
--source sa1b \
--sa1b-index-file /path/to/sa-1b.txt
(Optional) Step 4: Verify the reconstructed real-image directory
Run the downloader in dry-run mode against the same directory:
python download_real_images.py \
--metadata real_image_metadata.csv \
--output-dir ./real_images \
--sa1b-index-file /path/to/sa-1b.txt \
--dry-run
If the reconstruction is complete, the dry run should report that no real images still need to be downloaded or extracted.
License & Usage Terms
This repository is a curated subset for research reproducibility.
Repository Content: This dataset contains images sampled from third-party sources (COCO, LAION-2B, CC3M). We provide this gated access solely for non-commercial academic research and "Fair Use" computational analysis.
Original Licenses:
- COCO: Images are subject to the CC-BY 4.0 License.
- LAION / CC3M: These images are subject to the original licenses of their respective copyright holders.
Your Responsibility: By downloading this data, you agree to comply with the original licenses of the source datasets. The authors of X-AIGD do not claim ownership of the images provided in this repo and are not liable for any use outside of the agreed-upon research scope.
- Downloads last month
- 13