buxtcodes commited on
Commit
ed10612
·
verified ·
1 Parent(s): 3446b75

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +106 -0
README.md ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ task_categories:
4
+ - image-classification
5
+ tags:
6
+ - ai-generated-image-detection
7
+ - deepfake-detection
8
+ - wildfake
9
+ ---
10
+
11
+ # WildFake-Sample
12
+
13
+ A 30,000-image curated sample drawn from **WildFake** (Hao et al., AAAI 2025,
14
+ [arXiv:2402.11843](https://arxiv.org/abs/2402.11843);
15
+ [original dataset](https://modelscope.cn/datasets/hy2628982280/WildFake)),
16
+ built for evaluating AI-generated-image detectors' generalization to
17
+ generators and real-image sources they were **not** trained on. Built for
18
+ [`Buxt-Codes/AIGI-Detection`](https://github.com/Buxt-Codes/AIGI-Detection)
19
+ (branch `LoRC-PC`) — see that repo's `HANDOFF.md` for the full evaluation
20
+ methodology and results.
21
+
22
+ **This is a sample, not the full WildFake dataset** (which is ~3.6M images
23
+ across the same generators). All credit for the underlying images and the
24
+ dataset's design goes to WildFake's original authors — see their paper and
25
+ dataset page for licensing/usage terms; this sample exists purely to give
26
+ a fixed, reproducible evaluation slice across machines/checkpoints without
27
+ re-downloading from the ~39 multi-GB source archives every time.
28
+
29
+ ## Why this composition
30
+
31
+ The detector this sample was built to evaluate trained on SD-based fakes
32
+ over COCO real images (DDA-Training-Set) and was previously evaluated on
33
+ FLUX fakes (SID). This sample deliberately covers generators and real
34
+ sources **outside** both of those: 26 fake generators spanning GAN-based,
35
+ diffusion-based (non-SD and SD-family), and other architectures, plus 6 real
36
+ sources. WildFake's own COCO real split is deliberately excluded here since
37
+ the detector's training data is COCO-based — including it would risk
38
+ sampling images already seen in training.
39
+
40
+ ## Composition
41
+
42
+ **Fake — 750 images per generator (19,500 total):**
43
+
44
+ | Category | Generators |
45
+ |---|---|
46
+ | GAN-based | BigGAN, StyleGAN, StarGAN, DF-GAN, GALIP, GigaGAN |
47
+ | Diffusion (non-SD) | ADM, DDPM, DDIM, Imagen, VQDM, DALL-E 2, DALL-E 3, Midjourney v4, Midjourney v5 |
48
+ | Diffusion (SD family) | SDXL, Original SD, SD+ControlNet, SD+LoRA, SD+LyCORIS, Personalized SD (DreamBooth), Personalized SD (finetune) |
49
+ | Other | MAGE, VQGAN, VQVAE, MAE |
50
+
51
+ **Real — 1,750 images per source (10,500 total):** LAION-5B, ImageNet,
52
+ LSUN-Church, FFHQ, AFHQ, CelebA-HQ.
53
+
54
+ Total: **30,000 images** (19,500 fake / 10,500 real).
55
+
56
+ ## Files
57
+
58
+ ```
59
+ real/<source>/*.png 10,500 real images, 6 sources
60
+ fake/<generator>/*.png 19,500 fake images, 26 generators
61
+ manifest.csv / .json one row per image: which archive/path it came
62
+ from inside WildFake, category, split, resolution
63
+ transform_plan.csv / .json a fixed, deterministic per-image assignment of
64
+ one of 14 corruption conditions (JPEG/blur/resize/
65
+ noise/color-jitter/center-crop at various
66
+ severities) — see below
67
+ ```
68
+
69
+ Every image is a PNG regardless of its original format inside WildFake (many
70
+ of the fakes are natively PNG; reals are often JPEG) — deliberately not
71
+ JPEG-recompressed at sampling time, so any JPEG compression applied during
72
+ evaluation is the only lossy step, applied once, consistently.
73
+
74
+ ### `manifest.csv` columns
75
+ `split` (real/fake), `group` (generator or real-source name), `category`
76
+ (GAN_based/Diffusion_based/Other_based/Real), `source_zip` (which WildFake
77
+ archive), `source_path` (path inside that archive), `local_path` (path in
78
+ this dataset), `width`, `height`.
79
+
80
+ ### `transform_plan.csv` columns
81
+ `local_path`, `group`, `split`, `category`, `width`, `height`, `area`
82
+ (width×height), `condition` (e.g. "JPEG q=30", "Blur sigma=1.0"), `family`
83
+ (JPEG/Blur/Resize/Noise/ColorJitter/CenterCrop), `order` (always
84
+ `transform_first` — the condition is applied to the image, then a final
85
+ standard q=96 JPEG pass is applied last, standardizing the final encoding
86
+ every image ends up in regardless of which condition it drew).
87
+
88
+ **How the assignment was made** (no randomness, fully reproducible from
89
+ `manifest.csv` alone): within each group independently, images are sorted
90
+ by pixel area ascending, then cycled round-robin through the 14 conditions.
91
+ Sorting before cycling stratifies the assignment across resolution (any 14
92
+ consecutive area-sorted images cover every condition once, so no condition
93
+ is systematically biased toward small or large images); cycling makes it
94
+ uniform (every condition gets within 1 image of an even split per group).
95
+
96
+ ## How this was built
97
+
98
+ Sampled from WildFake's ModelScope-hosted archives (several 51–53GB each,
99
+ some bundling multiple generators together) via HTTP Range requests —
100
+ reading only each archive's central directory (a listing) plus the exact
101
+ compressed bytes of the specific images sampled, never downloading an
102
+ archive in full. First-N selection per generator/source in each archive's
103
+ own listing order (no random sampling). See
104
+ `data_prep/build_wildfake_eval.py` and
105
+ `data_prep/build_wildfake_transform_plan.py` in the GitHub repo above for
106
+ the exact code.