Shanmuk4622/ai-image-detection-dataset

AI-Image Detection Dataset

Paired real / AI images, with shared image-grounded captions, for training and evaluating AI-generated-image detectors.

Each of 10,000 real photos is captioned once (BLIP-2) and paired with one synthetic partner per generator (6 generators → 60,000 AI images). A real image and all of its AI partners share the same prompt, so the only systematic difference between the classes is the generative process itself. A detector trained here is pushed toward the synthesis fingerprint (texture, frequency, rendering artefacts) instead of scene content — the shortcut that inflates accuracy on naively-built datasets.

This is the clean, self-contained release. Every row is complete on its own: the image renders in the viewer, carries its prompt, its train/val/test split, and full caption + generation provenance — no need to join against side files. The viewer exposes a default subset (all 70k) plus one subset per model, each with its own train/validation/test splits.

At a glance

Real images	10,000 (COCO + ImageNet-1k), label 0
AI images	60,000 across 6 generators, label 1
Total rows	70,000
Pairs	10,000 (each real ↔ 6 AI partners, shared caption)
Resolution	512×512 RGB PNG
Splits	train 49,392 / validation 10,122 / test 10,486 (pair-level, seed 42)
Subsets	default + real, sd15, sdxl, flux_schnell, kandinsky22, pixart_sigma, wuerstchen
Preprocessing	Identical canonical pipeline for real and AI (pipeline_version 1.2)
License	Non-commercial research use (most-restrictive inherited term)

Subsets

The Dataset Viewer and load_dataset expose eight subsets, each split into train / validation / test:

subset	rows	contents
default	70,000	everything (real + all generators)
real	10,000	real photos only (label 0)
sd15	10,000	Stable Diffusion 1.5 outputs
sdxl	10,000	SDXL outputs
flux_schnell	10,000	FLUX.1-schnell outputs
kandinsky22	10,000	Kandinsky 2.2 outputs
pixart_sigma	10,000	PixArt-Σ outputs
wuerstchen	10,000	Würstchen outputs

What changed vs. the earlier release

• image is now the Hugging Face Image type → thumbnails render in the Dataset Viewer.
• prompt is populated on real rows (each real image's own BLIP-2 caption), so a pair literally shares one caption.
• split is a column and the files are organised into <subset>/{train,val,test}/, so load_dataset(..., split=...) and the viewer's split tabs work.
• Per-model subsets are selectable directly in the viewer / loader.
• Caption provenance added: raw_caption, caption_n_tokens, caption_model, plus the original BLIP-2 caption shards under captions/.

Generators

key	model id	native res	steps	guidance
sd15	stable-diffusion-v1-5/stable-diffusion-v1-5	512	20	7.0
sdxl	stabilityai/stable-diffusion-xl-base-1.0	1024	8	7.0
flux_schnell	black-forest-labs/FLUX.1-schnell	1024	4	0.0
kandinsky22	kandinsky-community/kandinsky-2-2-decoder	768	25	4.0
pixart_sigma	PixArt-alpha/PixArt-Sigma-XL-2-1024-MS	1024	20	4.5
wuerstchen	warp-ai/wuerstchen	1024	[20, 10]	[4.0, 0.0]

How the pairs are built

Real images are captioned with Salesforce/blip2-opt-2.7b, the caption is cleaned and capped to 75 CLIP tokens, and that exact caption is reused by all six generators and also stored on the real row. source_real_id links every AI image back to its real partner (source_real_id == image_id for real rows).

Canonical preprocess (leak-free)

Both real and AI images pass through the same pipeline:

1. EXIF-transpose → convert to RGB → center-crop to square → Lanczos resize to 512.
2. JPEG-equalise (quality 95, 4:4:4) → save as PNG (compress level 6, no EXIF/ICC).

Applying an identical transform to both classes removes the resolution, colour-space and JPEG-artefact shortcuts that would otherwise let a model cheat.

Splits

Deterministic pair-level split keyed on source_real_id (seed 42): a real image and all six of its AI partners always land in the same fold, so there is no scene leakage between train / validation / test.

split	images
train	49,392
validation	10,122
test	10,486

Dataset structure

ai-image-detection-dataset/
├── real/            {train,val,test}/*.parquet    # 10k real photos, label 0
├── sd15/            {train,val,test}/*.parquet    # 10k AI per generator, label 1
├── sdxl/            {train,val,test}/*.parquet
├── flux_schnell/    {train,val,test}/*.parquet
├── kandinsky22/     {train,val,test}/*.parquet
├── pixart_sigma/    {train,val,test}/*.parquet
├── wuerstchen/      {train,val,test}/*.parquet
├── captions/        captions-*.parquet            # original BLIP-2 caption shards (provenance)
├── metadata/
│   ├── manifest.parquet        # id, pairing key, label, generator, split (no image bytes)
│   ├── splits.parquet          # source_real_id -> split
│   ├── captions.parquet        # consolidated real-image captions + provenance
│   ├── config.json             # build configuration (seeds, models, params)
│   └── validation_report.json  # leak-audit / sniff-test results
├── ai_detect_common.py         # shared preprocess + schema library
└── README.md

Usage

from datasets import load_dataset

# everything, by split
ds = load_dataset("Shanmuk4622/ai-image-detection-dataset")          # default subset
train, val, test = ds["train"], ds["validation"], ds["test"]

ex = train[0]
ex["image"]     # PIL.Image (decoded automatically, renders in the viewer)
ex["label"]     # 0 = real, 1 = AI
ex["prompt"]    # shared caption (now present on real rows too)
ex["generator"] # 'real' or one of the 6 generators
ex["split"]     # 'train' | 'validation' | 'test'

# a single model's subset (real + one generator are separate subsets)
flux_test = load_dataset("Shanmuk4622/ai-image-detection-dataset", "flux_schnell", split="test")
real_train = load_dataset("Shanmuk4622/ai-image-detection-dataset", "real", split="train")

Schema

column	type	description
image	Image struct{bytes, path}	canonical 512×512 RGB PNG
image_id	string	globally unique, e.g. real_000001 / sdxl_000001
source_real_id	string	pairing key; equals image_id for real rows
label	int8	0 = real, 1 = AI
generator	string	real, sd15, sdxl, flux_schnell, kandinsky22, pixart_sigma, wuerstchen
source_dataset	string	coco, imagenet, or <generator>
split	string	train / validation / test
prompt	string	shared image-grounded caption (now populated for real rows)
raw_caption	string	uncleaned BLIP-2 output
caption_n_tokens	int16	CLIP-token length of prompt
caption_model	string	captioner used (Salesforce/blip2-opt-2.7b)
width / height	int16	always 512
orig_width / orig_height	int32	provenance only — never use as a training feature
gen_model_id	string	HF model id (null for real)
gen_steps / gen_guidance / gen_native_res	int16 / float32 / int16	generation params (null for real)
seed	int64	hash(master_seed, generator, source_real_id) (null for real)
pipeline_version	string	1.2
sha256	string	hex digest of the PNG bytes
created_utc	string	ISO-8601 UTC timestamp

Leak audit

A "sniff test" trains a lightweight classifier on each generator vs. real using only low-level statistics and checks separability stays healthy (<0.85 healthy, 0.85–0.95 investigate, >0.95 leak):

generator	score
sd15	0.910
sdxl	0.973
flux_schnell	0.917
kandinsky22	0.958
pixart_sigma	0.943
wuerstchen	0.843
overall	0.860

Full results in metadata/validation_report.json.

Provenance, license and intended use

• ImageNet content is non-commercial research only (ILSVRC terms).
• COCO images are Flickr-sourced (Creative Commons).
• AI images are synthetic outputs of the models listed above, each under its own license.
• This dataset inherits the most restrictive applicable term: non-commercial research use. Not legal advice.

Limitations and caveats

• Core set is text-to-image only (no img2img / reference conditioning); detectors trained here target the text-to-image threat model.
• Caption quality (BLIP-2 concise sentences) bounds prompt diversity.
• FLUX.1-schnell and Würstchen run with CPU offload on T4; generation conditions are otherwise consistent with standard inference settings.
• Per-model step counts were reduced for speed (SDXL 8 steps, SD 1.5 20 steps), which is representative of real-world fast inference.

Citation

@misc{aiimage_detection_dataset,
  title  = {AI-Image Detection Dataset},
  author = {Shanmuk4622},
  year   = {2025},
  howpublished = {\url{https://huggingface.co/datasets/Shanmuk4622/ai-image-detection-dataset}}
}