Zitacron/real-vs-ai-corpus

Real vs AI Corpus

Upload in progress. This dataset is still being built — new shards are added continuously. The default config (used by the dataset viewer) is a 1 000-row proportional seed sample that reflects the current label distribution accurately. Use name="full" in code to load the complete, growing dataset.

A large-scale binary image classification dataset for training AI-image detectors. Built by Zitacron from 17 public HuggingFace sources, all streaming-merged with no intermediate local storage.

All constituent sources are CC BY 4.0, Apache 2.0, or MIT — fully commercially usable. Models trained on this dataset may be used commercially without restriction, provided attribution requirements below are met.

---

Schema

Column	Type	Description
image	Image	PIL image (RGB)
label	ClassLabel	0 = real · 1 = ai
label_text	string	"real" or "ai"
source_dataset	string	HuggingFace repo ID of the origin dataset
source_license	string	SPDX license ID of the origin dataset

---

Sources & Licences

AI-generated images

Dataset	Rows (approx.)	License	Attribution required
svjack/diffusiondb_random_10k	10 k	CC BY 4.0	Cite DiffusionDB (Wang et al., 2022)
bitmind/nano-banana	9.5 k	MIT	Keep MIT copyright notice
ash12321/nano-banana-pro-generated-1k	1 k	MIT	Keep MIT copyright notice
julienlucas/midjourney-dalle-sd-nanobananapro-dataset	11 k (mixed)	CC BY 4.0	Attribute julienlucas
AbstractPhil/synthetic-characters ×3 configs	~150 k	CC BY 4.0	Attribute AbstractPhil
LucasFang/FLUX-Reason-6M	5.9 M	Apache 2.0	Include NOTICE / attribution in derivatives
Parveshiiii/AI-vs-Real	14 k (mixed)	CC BY 4.0	Attribute Parveshiiii

Real images

Dataset	Rows (approx.)	License	Attribution required
ronantakizawa/webui	29 k	CC BY 4.0	Attribute ronantakizawa
derek-thomas/ScienceQA	6 k	CC BY 4.0	Cite ScienceQA (Lu et al., 2022)
skylenage/DeepVision-103K (visual_logic config)	26 k	CC BY 4.0	Attribute skylenage
MBZUAI/OpenEarthAgent	11.6 k	CC BY 4.0	Attribute MBZUAI
EPFL-ECEO/coralscapes	1.5 k	CC BY 4.0	Attribute EPFL-ECEO
opendatalab/OmniDocBench	1.4 k	Apache 2.0	Include NOTICE / attribution in derivatives
Sigurdur/isl-finepdfs-images	1.1 k	CC BY 4.0	Attribute Sigurdur
laion/laion2B-en-aesthetic (URL stream, capped for balance)	up to 5.92 M (actual varies — dead URLs skipped)	CC BY 4.0	Cite LAION-5B (Schuhmann et al., 2022)

---

Combined licence

CC BY 4.0 — the combined dataset is released under the most restrictive licence present in the sources. Individual rows carry their source_license field for granular provenance.

Permitted Uses

• Use this dataset to train, fine-tune, or evaluate any model.
• Deploy models trained on it commercially.
• Redistribute this dataset or derivatives.
• Build products on top of models trained with it.

Attribution Requirements

• Attribute this dataset as:

  Real vs AI Corpus, Zitacron, 2026.
  https://huggingface.co/datasets/Zitacron/real-vs-ai-corpus

• Retain the source_dataset and source_license columns if redistributing subsets, so downstream users can trace each image to its origin.
• For Apache 2.0 sources (LucasFang/FLUX-Reason-6M, opendatalab/OmniDocBench): if you redistribute a derivative dataset that includes only those rows, also include their upstream NOTICE file if one exists.
• For MIT sources (bitmind/nano-banana, ash12321/nano-banana-pro-generated-1k): keep the upstream copyright notice in any redistribution of those rows.

Additional Permissions (No Copyleft)

• You are not required to open-source models trained on this dataset.
• You are not required to share-alike (this is not a copyleft licence).
• Model weights trained on this dataset are unrestricted — you may keep them proprietary, sell them, or publish them under any licence you choose.

---

Citation

If you use this dataset, please cite:

@dataset{zitacron_realvsai_2026,
  author    = {zitacron},
  title     = {Real vs AI Corpus},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/Zitacron/real-vs-ai-corpus}
}

And the upstream sources your rows originate from (see table above).

---

Usage

from datasets import load_dataset

# Full dataset — streaming (recommended; dataset is large and still growing)
ds = load_dataset("Zitacron/real-vs-ai-corpus", name="full", streaming=True, split="train")
for row in ds:
    img        = row["image"]          # PIL Image (RGB)
    label      = row["label"]          # int  — 0 = real, 1 = ai
    label_text = row["label_text"]     # str  — "real" or "ai"
    source     = row["source_dataset"] # str  — origin HF repo
    lic        = row["source_license"] # str  — SPDX id

# 1 000-row seed sample (fast, accurate label distribution preview)
ds_seed = load_dataset("Zitacron/real-vs-ai-corpus", name="default", split="train")
print(ds_seed[0])

Filter to a single source

from datasets import load_dataset

ds = load_dataset("Zitacron/real-vs-ai-corpus", name="full", streaming=True, split="train")
flux_only = ds.filter(lambda r: r["source_dataset"] == "LucasFang/FLUX-Reason-6M")

Balance classes

import itertools
from datasets import load_dataset

ds   = load_dataset("Zitacron/real-vs-ai-corpus", streaming=True, split="train")
real = ds.filter(lambda r: r["label"] == 0)
ai   = ds.filter(lambda r: r["label"] == 1)
# Round-robin interleave for balanced batches
balanced = real.interleave(ai)  # or use datasets.interleave_datasets