KarmaLeo/Xiz9Dataset

Deepfake / Synthetic Image Detection Dataset (built for an undergraduate thesis)

This dataset combines a subset of AI-Generated Images vs Real Images (by tristanzhang32, on Kaggle) with an extracted subset of OpenFake (by ComplexDataLab), assembled to train and evaluate a deepfake/synthetic-image detector for an undergraduate thesis (TCC) at Instituto Mauá de Tecnologia (IMT).

This is an academic project, not a commercial product, developed as part of the requirements for an undergraduate degree at Instituto Mauá de Tecnologia.

This is not an original dataset. A portion of it was collected independently (see below); the rest is a filtered subset of OpenFake v2. No new images were generated by the author. This repository only selects, reorganizes, and merges existing data.

Why it's built this way

The starting point was AI-Generated Images vs Real Images from Kaggle:

• Fake: 10,000 Stable Diffusion, 10,000 MidJourney, 10,000 DALL-E
• Real: 22,500 Pexels/Unsplash, 7,500 WikiArt

To grow this dataset without duplicating the same generators/sources, an additional subset was extracted from OpenFake, deliberately excluding Stable Diffusion, MidJourney, and DALL-E outputs (already covered above), and adding only generators/sources not already present.

Composition

Category	Original	Added from OpenFake	Total
Fake	30,000	53,555	83,555
Real	30,000	53,555	83,555
Total	60,000	107,110	167,110

Fake sources

Original (10,000 each): Stable Diffusion, MidJourney, DALL-E

Added from OpenFake (53,555 total, 50+ generators), main contributors:

Generator	Images
flux.1-dev	10,000
gpt-image-1	5,019
flux.1-schnell	4,837
flux-1.1-pro	3,354
hidream-i1-full	3,314
ideogram-3.0	3,261
flux-mvc5000	1,859
mystic	1,819
flux.2-dev	1,770
imagen-4.0	1,630
grok-2-image-1212	1,160
...	40+ additional generators, community fine-tunes, and LoRAs

The counts above reflect the generators as extracted; a per-generator breakdown is included for transparency, but individual images are not labeled by generator in the final folder structure.

Real sources

Original: Pexels/Unsplash (22,500), WikiArt (7,500)

Added from OpenFake:

Source	Images
LAION	27,318
Pexels	26,237

Structure

data/
  ├── train-00000.parquet
  ├── train-00001.parquet
  ├── ...
  ├── test-00000.parquet
  ├── test-00001.parquet
  └── ...

Data is stored as Parquet shards (not loose image files), following the standard Hugging Face datasets convention. Each row has:

• image: the image, stored as raw bytes (decoded automatically as an image by datasets/the Hub viewer)
• label: fake or real
• split: train or test

Loading with datasets will decode the image column automatically:

from datasets import load_dataset
ds = load_dataset("KarmaLeo/Xiz9Dataset")

The train/test split (roughly 80/20) mirrors the original dataset's proportions; new images were assigned to train/test at random while preserving that ratio, and checked against the existing set by content hash to avoid duplicates.

License

This dataset does not have a single uniform license. It follows a mixed license, matching each portion's original terms:

• CC-BY-SA-4.0 for the dataset as a whole (attribution + share-alike).
• Non-commercial use required for images produced by proprietary generators (gpt-image-1, imagen-4.0, imagen-3.0-002, ideogram-3.0, grok-2-image-1212, grok-imagine, nano-banana, nano-banana-2, seedream-v4.0, seedream-v4.5, mystic, kling-video-v2.5-turbo, and others), due to non-compete clauses from those providers. This restriction is inherited directly from OpenFake and has not been changed or removed here.
• Stable Diffusion, MidJourney, and DALL-E outputs, and the Pexels/Unsplash/WikiArt real images, come from AI-Generated Images vs Real Images on Kaggle, released under the MIT License.

For the full per-source/per-generator licensing breakdown, see the original OpenFake dataset and its paper (ethics and licensing section).

Because part of this dataset is restricted to non-commercial use, treat the dataset as a whole as non-commercial, even though the MIT-licensed portion would otherwise permit commercial use on its own.

This dataset is released for academic and research purposes (undergraduate thesis), not for commercial use.

Citation

If you use the Kaggle-derived portion of this dataset, please reference the source:

AI-Generated Images vs Real Images
https://www.kaggle.com/datasets/tristanzhang32/ai-generated-images-vs-real-images

If you use the OpenFake-derived portion of this dataset, please cite the original work:

OpenFake: An Open Dataset and Platform Toward Real-World Deepfake Detection
arXiv:2509.09495
https://huggingface.co/datasets/ComplexDataLab/OpenFake

Disclaimer

This repository is not affiliated with, endorsed by, or maintained by ComplexDataLab or by the creator of the Kaggle source dataset. It is an independent compilation assembled as part of an undergraduate thesis (TCC) project at Instituto Mauá de Tecnologia (IMT).