WildFake: A Large-scale Challenging Dataset for AI-Generated Images Detection
Paper • 2402.11843 • Published
Datasets:
image_bytes unknown | split stringclasses 1
value | group stringclasses 16
values | category stringclasses 2
values | source_zip stringclasses 10
values | source_path stringlengths 22 97 | width int32 200 2.69k | height int32 192 3.07k | condition stringclasses 14
values | family stringclasses 6
values | order stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|---|
[
137,
80,
78,
71,
13,
10,
26,
10,
0,
0,
0,
13,
73,
72,
68,
82,
0,
0,
0,
200,
0,
0,
0,
200,
8,
2,
0,
0,
0,
34,
58,
57,
201,
0,
0,
115,
24,
73,
68,
65,
84,
120,
156,
221,
253,
233,
154,
36,
57,
146,
45,
136,
29,
1... | fake | BigGAN | GAN_based | Images/GAN_based.zip | GAN_based/Typical/BigGAN/574/img005288.jpg | 200 | 200 | JPEG q=90 | JPEG | transform_first |
[
137,
80,
78,
71,
13,
10,
26,
10,
0,
0,
0,
13,
73,
72,
68,
82,
0,
0,
0,
200,
0,
0,
0,
200,
8,
2,
0,
0,
0,
34,
58,
57,
201,
0,
0,
192,
168,
73,
68,
65,
84,
120,
156,
196,
253,
215,
150,
36,
73,
114,
40,
8,
138,
1... | fake | BigGAN | GAN_based | Images/GAN_based.zip | GAN_based/Typical/BigGAN/574/img005289.jpg | 200 | 200 | JPEG q=70 | JPEG | transform_first |
[
137,
80,
78,
71,
13,
10,
26,
10,
0,
0,
0,
13,
73,
72,
68,
82,
0,
0,
0,
200,
0,
0,
0,
200,
8,
2,
0,
0,
0,
34,
58,
57,
201,
0,
1,
0,
0,
73,
68,
65,
84,
120,
156,
92,
253,
221,
118,
28,
71,
178,
52,
136,
154,
185,... | fake | BigGAN | GAN_based | Images/GAN_based.zip | GAN_based/Typical/BigGAN/574/img005281.jpg | 200 | 200 | JPEG q=50 | JPEG | transform_first |
"iVBORw0KGgoAAAANSUhEUgAAAMgAAADICAIAAAAiOjnJAACGXElEQVR4nLX9aZMrSXYlCJ5z1QyAu781llwjSRa3aulpmZkWmS/(...TRUNCATED) | fake | BigGAN | GAN_based | Images/GAN_based.zip | GAN_based/Typical/BigGAN/574/img005282.jpg | 200 | 200 | JPEG q=30 | JPEG | transform_first |
"iVBORw0KGgoAAAANSUhEUgAAAMgAAADICAIAAAAiOjnJAAEAAElEQVR4nDT92ZJkOZIlCDIYO+4mIrqYmZt7eEROZndTTVHP58x(...TRUNCATED) | fake | BigGAN | GAN_based | Images/GAN_based.zip | GAN_based/Typical/BigGAN/574/img005284.jpg | 200 | 200 | Blur sigma=0.5 | Blur | transform_first |
"iVBORw0KGgoAAAANSUhEUgAAAMgAAADICAIAAAAiOjnJAADPYUlEQVR4nJT9WZPkSJImCH7MLAKoqpkfER6RV1XW0d3V1TPVD7O(...TRUNCATED) | fake | BigGAN | GAN_based | Images/GAN_based.zip | GAN_based/Typical/BigGAN/574/img005283.jpg | 200 | 200 | Blur sigma=1.0 | Blur | transform_first |
"iVBORw0KGgoAAAANSUhEUgAAAMgAAADICAIAAAAiOjnJAACGzUlEQVR4nJX9za4sSZImBn6fqLqfcyMiK7uaXdUcoguNIYFeDMD(...TRUNCATED) | fake | BigGAN | GAN_based | Images/GAN_based.zip | GAN_based/Typical/BigGAN/574/img005286.jpg | 200 | 200 | Blur sigma=2.0 | Blur | transform_first |
"iVBORw0KGgoAAAANSUhEUgAAAMgAAADICAIAAAAiOjnJAADnw0lEQVR4nHT92ZIlSZIlBh5mFlHVe818iYhcInKrzKpEd1UWgG7(...TRUNCATED) | fake | BigGAN | GAN_based | Images/GAN_based.zip | GAN_based/Typical/BigGAN/574/img005287.jpg | 200 | 200 | Resize 0.5x | Resize | transform_first |
"iVBORw0KGgoAAAANSUhEUgAAAMgAAADICAIAAAAiOjnJAADkRElEQVR4nGT955IkSXItCKtxJ0GSFG0yDMBiV773f4j7DCt3gcX(...TRUNCATED) | fake | BigGAN | GAN_based | Images/GAN_based.zip | GAN_based/Typical/BigGAN/574/img005285.jpg | 200 | 200 | Resize 0.25x | Resize | transform_first |
"iVBORw0KGgoAAAANSUhEUgAAAMgAAADICAIAAAAiOjnJAAEAAElEQVR4nEz96Y4kWZIuiMlyNl1scQ+PiKzM6r5d0zP3zgzBh+D(...TRUNCATED) | fake | BigGAN | GAN_based | Images/GAN_based.zip | GAN_based/Typical/BigGAN/574/img005280.jpg | 200 | 200 | Noise sigma=0.02 | Noise | transform_first |
End of preview. Expand in Data Studio
WildFake-Sample
A 30,000-image sample of WildFake (Hao et al., AAAI 2025,
arXiv:2402.11843;
original dataset),
covering generators and real-image sources outside DDA/SID — a held-out
generalization slice, not a copy of the full ~3.6M-image dataset. All credit
for the images goes to WildFake's original authors. Built for
Buxt-Codes/AIGI-Detection
(branch LoRC-PC) — see that repo's HANDOFF.md for the evaluation
methodology and results.
Composition
- Fake (19,500): 750 × 26 generators — GANs (BigGAN, StyleGAN, StarGAN, DF-GAN, GALIP, GigaGAN), non-SD diffusion (ADM, DDPM, DDIM, Imagen, VQDM, DALL-E 2/3, Midjourney v4/v5), SD-family (SDXL, OriginalSD, ControlNet, LoRA, LyCORIS, 2x personalized), other (MAGE, VQGAN, VQVAE, MAE).
- Real (10,500): 1,750 × 6 sources — LAION-5B, ImageNet, LSUN-Church, FFHQ, AFHQ, CelebA-HQ.
Files
data/train-*.parquet — one row per image. manifest.csv/.json and
transform_plan.csv/.json — the same metadata as plain CSV/JSON.
Columns: image_bytes (raw file bytes, undecoded — decode with
Image.open(io.BytesIO(row["image_bytes"]))), split, group, category,
source_zip/source_path, width/height, condition (one of 14
transform-battery conditions, assigned round-robin per group, stratified by
resolution), family, order (always transform_first: condition applied,
then a final standardizing JPEG q=96 pass).
Full build methodology (HTTP range-request sampling, transform-assignment algorithm, source code): see the GitHub repo linked above.
- Downloads last month
- 6