Datasets:

nebula
/

OpenSDIDplus

pretty_name: OpenSDID+
license: cc-by-sa-4.0
task_categories:
  - image-classification
  - image-segmentation
size_categories:
  - 100K<n<1M
tags:
  - diffusion
  - image-generation
  - ai-generated-content
  - deepfake-detection
  - synthetic-image-detection
  - image-forensics
  - localization
  - opensdi
  - arxiv:2503.19653
dataset_info:
  features:
    - name: key
      dtype: string
    - name: image
      dtype: image
    - name: mask
      dtype: image
    - name: label
      dtype:
        class_label:
          names:
            '0': real
            '1': fake
  splits:
    - name: sd3
      num_bytes: 48619528217.1
      num_examples: 211676
    - name: sd2
      num_bytes: 29114150644.648
      num_examples: 182196
    - name: sdxl
      num_bytes: 41797503707.276
      num_examples: 200302
    - name: flux
      num_bytes: 35912899091.312
      num_examples: 208344
  download_size: 144545739362
  dataset_size: 155444081660.336
configs:
  - config_name: default
    data_files:
      - split: sd3
        path: data/sd3-*
      - split: sd2
        path: data/sd2-*
      - split: sdxl
        path: data/sdxl-*
      - split: flux
        path: data/flux-*

OpenSDID+

OpenSDID+ is an extended release of the OpenSDI dataset. It complements the original SD1.5 training split with large-scale images from the remaining OpenSDI generators: SD2, SD3, SDXL, and FLUX.

The dataset follows the OpenSDI challenge introduced in "OpenSDI: Spotting Diffusion-Generated Images in the Open World". OpenSDI studies detection and localization of diffusion-generated images under realistic open-world settings, including diverse user intentions, evolving diffusion models, and both global and local image manipulations.

Relationship to OpenSDI

The original OpenSDI release is split across:

OpenSDI_train: SD1.5 training data
OpenSDI_test: benchmark test data across OpenSDI generators

This repository provides the expanded generator-specific data for SD2, SD3, SDXL, and FLUX. It is intended for researchers who want broader generator coverage beyond the original SD1.5 training split.

Dataset Details

Split	Generator family	Samples
`sd2`	Stable Diffusion 2.x	182,196
`sd3`	Stable Diffusion 3	211,676
`sdxl`	Stable Diffusion XL	200,302
`flux`	FLUX	208,344
Total		802,518

Each example contains:

key: sample identifier
image: input image
mask: manipulation/localization mask
label: binary class label, where 0 = real and 1 = fake

Supported Tasks

Diffusion-Generated Image Detection

Classify whether an image is real or diffusion-generated.

Manipulation Localization

Use the mask annotations to evaluate or train pixel-level localization methods for locally manipulated images.

Usage

from datasets import load_dataset

# Choose one generator split: "sd2", "sd3", "sdxl", or "flux".
ds = load_dataset("nebula/OpenSDIDplus", split="sdxl", streaming=True)
sample = next(iter(ds))
print(sample.keys())
print(sample["label"])

For full-scale training, streaming or selective split download is recommended because the release is large.

Intended Uses

OpenSDID+ is intended for academic research on:

AI-generated image detection
diffusion-generated image detection
image manipulation localization
open-world image forensics
generalization across unseen or evolving text-to-image generators

Limitations

OpenSDID+ should be interpreted as an extension of the OpenSDI benchmark rather than a standalone replacement for the original release. The splits in this repository cover SD2, SD3, SDXL, and FLUX; use OpenSDI_train for the original SD1.5 training split.

The dataset reflects the generator families, prompts, editing procedures, and image sources used in OpenSDI. Performance on OpenSDID+ may not fully predict robustness to future generators, post-processing pipelines, social media compression, or real-world distribution shifts.

Ethical Considerations

This dataset is released for research on synthetic media detection and image forensics. Users should not use it to create deceptive content, identify private individuals, or deploy forensic systems without appropriate validation, fairness analysis, and legal review.

Some generated or source images may contain sensitive or biased content inherited from public image and generative model ecosystems. Users should apply appropriate filtering and access controls where needed.

Citation

If you use OpenSDID+ or the OpenSDI benchmark, please cite:

@InProceedings{Wang_2025_CVPR,
  author    = {Wang, Yabin and Huang, Zhiwu and Hong, Xiaopeng},
  title     = {OpenSDI: Spotting Diffusion-Generated Images in the Open World},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2025},
  pages     = {4291--4301}
}