ekyleyang/DOCO-LAION-C

DOCO ImageNet-L and OOD-L with LAION-C Distortions

This repository contains the materialized ImageNet-L and OOD-L image sets used for the LAION-C experiments in Back to Source: Open-Set Continual Test-Time Adaptation via Domain Compensation (DOCO) (CVF PDF, arXiv:2604.21772, code).

The artifacts were generated by applying a modified version of the LAION-C corruption pipeline to ImageNet validation images and six OOD source datasets. The original LAION-C benchmark release is available separately on Zenodo (DOI: 10.5281/zenodo.14051887).

This release provides the exact generated images used by the DOCO evaluation code, including both severity 1 and severity 3. It is intended to make the reported experiments reproducible without regenerating stochastic corruptions.

Visual examples

Examples of the six corruption types included in this release. The top row shows severity 1 and the bottom row shows severity 3.

Released data

The repository contains 14 uncompressed tar archives: one archive for each of seven datasets at each of two severity levels.

Released directory	Role	Base images per corruption and severity	Severities
ImageNet-LAION-5K	ID	5,000	1, 3
Places365-L-6k	OOD	6,000	1, 3
Textures-L-6k	OOD	5,640	1, 3
iNaturalist-L-6k	OOD	6,000	1, 3
SUN-L-6k	OOD	6,000	1, 3
NINCO_OOD_classes-L-6k	OOD	5,399	1, 3
SSB-Hard-L-6k	OOD	6,000	1, 3

Each image set contains six LAION-C corruption types:

• mosaic
• vertical_lines
• glitched
• luminance
• geometric_shapes
• sticker

Across both severities, the release contains 480,468 JPEG files. The 14 archives total 51,459,164,160 bytes (47.93 GiB). The directory names ending in -6k are retained for compatibility with the DOCO code; Textures and NINCO contain all supported source images available to the generator, which is fewer than 6,000.

Source data and provenance

The generated images retain content from the following source datasets. This release does not replace the source datasets or their documentation.

DOCO output	Upstream source used for generation	Selection used in this release
ImageNet-LAION-5K	ImageNet ILSVRC2012 validation archive and development kit	The 5,000 validation images listed in imagenet_test_image_ids_5k.txt
Places365-L-6k	Places365 val_256.tar download page	First 6,000 supported files after sorting
Textures-L-6k	Describable Textures Dataset (DTD)	All 5,640 supported files; class folders are preserved
iNaturalist-L-6k	iNaturalist OOD archive, distributed for the MOS large-scale OOD benchmark	First 6,000 supported files after sorting
SUN-L-6k	SUN OOD archive, distributed for the MOS large-scale OOD benchmark	First 6,000 supported files after sorting
NINCO_OOD_classes-L-6k	NINCO OOD classes download folder, linked by COME	All 5,399 supported files; class folders are preserved
SSB-Hard-L-6k	SSB-Hard download folder, linked by COME	6,000 files sampled with NumPy RandomState seed 22; class folders are preserved

For the three flat source directories (Places365, iNaturalist, and SUN), the generator sorts supported files by name and takes the first 6,000. For nested datasets, it sorts class folders and filenames before selection. Textures and NINCO have fewer than 6,000 supported JPEG files, so every available file is used. SSB-Hard is sampled after sorting with seed 22.

The mosaic and sticker corruptions use a WebDataset tile archive created from ImageNet validation images by the released preprocessing code. The source ImageNet archives and the intermediate tile archive are not included as separate files here.

Generation code

The artifacts correspond to DOCO commit c2f4c8ba566ba38222a16c2bc18805962ac8d728.

The relevant scripts are:

1. prerun_create_webdataset.py, which builds the ImageNet validation tile archive used by mosaic and sticker.
2. run_laionc_synthetic_ImageNet.py, which generates ImageNet-L from the fixed ImageNet 5k validation subset.
3. run_laionc_synthetic_sixOOD.py, which generates the six OOD-L datasets.

The generator preprocesses the released images to 224 x 224. The evaluation path therefore avoids an additional resize and crop.

The archives and SHA256SUMS define the exact released artifacts. Although dataset selection is deterministic as described above, several NumPy-based stochastic corruption operations in the released generator do not share one global generation seed. Use these materialized archives when exact image-level reproduction is required.

Archive layout

Each archive extracts directly under a common dataset root and preserves the paths expected by DOCO:

<DATA_ROOT>/
  ImageNet-LAION-5K/<corruption>/intensity_level_<1-or-3>/<class>/*.JPEG
  Places365-L-6k/<corruption>/intensity_level_<1-or-3>/*.JPEG
  Textures-L-6k/<corruption>/intensity_level_<1-or-3>/<class>/*.JPEG
  iNaturalist-L-6k/<corruption>/intensity_level_<1-or-3>/*.JPEG
  SUN-L-6k/<corruption>/intensity_level_<1-or-3>/*.JPEG
  NINCO_OOD_classes-L-6k/<corruption>/intensity_level_<1-or-3>/<class>/*.JPEG
  SSB-Hard-L-6k/<corruption>/intensity_level_<1-or-3>/<class>/*.JPEG

Download, verify, and extract

Install the Hugging Face client and authenticate if this repository is private or gated:

python -m pip install -U huggingface_hub
hf auth login

Download severity 3, which is used for DOCO Table 2:

hf download ekyleyang/DOCO-LAION-C \
  --repo-type dataset \
  --include "archives/*/severity_3.tar" \
  --include SHA256SUMS \
  --local-dir ./DOCO-LAION-C-download

To download both released severities, replace the archive include pattern with archives/*/severity_*.tar.

Verify the downloaded archives and extract them:

cd ./DOCO-LAION-C-download
sha256sum -c SHA256SUMS --ignore-missing

export DATA_ROOT=/path/to/datasets
for archive in archives/*/severity_3.tar; do
  tar -xf "$archive" -C "$DATA_ROOT"
done

No Hugging Face dataset loader or code change is required. After extraction, the existing DOCO loaders read these folders directly.

DOCO Table 2 protocol

The released Table 2 launcher uses:

• severity 3
• batch size 64
• OOD ratio 0.5
• DOCO_BETA=0.5
• DOCO_PROMPT_NUM=8

See the DOCO repository for the complete environment and evaluation commands.

Terms and attribution

These archives contain transformed derivatives of images from ImageNet, Places365, DTD, iNaturalist, SUN, NINCO, and SSB-Hard. Each source retains its own license, access conditions, usage restrictions, and attribution requirements. This repository does not grant additional rights over the source images.

The Hugging Face metadata uses license: other because the source terms are mixed. Before downloading, redistributing, or using these artifacts, review the terms on every applicable upstream source page linked above. Please also follow the citation requirements of the source datasets used in your experiment.

Citation

If you use these released artifacts, please cite both DOCO and LAION-C. The DOCO citation below follows the official CVF Open Access entry. See also the arXiv record and DOCO code.

@InProceedings{Yang_2026_CVPR,
    author    = {Yang, Yingkai and Chen, Chaoqi and Huang, Hui},
    title     = {Back to Source: Open-Set Continual Test-Time Adaptation via Domain Compensation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {7957-7966}
}

@inproceedings{li2025laionc,
  title     = {LAION-C: An Out-of-Distribution Benchmark for Web-Scale Vision Models},
  author    = {Li, Fanfei and Klein, Thomas and Brendel, Wieland and Geirhos, Robert and Zimmermann, Roland S.},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning (ICML)},
  year      = {2025},
  url       = {https://arxiv.org/abs/2506.16950}
}

Acknowledgments

The corruption design and implementation are based on LAION-C. The iNaturalist and SUN distribution links come from MOS, and the NINCO and SSB-Hard distribution link comes from COME. We thank the dataset and code authors for making these research resources available.