Add dataset card
Browse files
README.md
CHANGED
|
@@ -1,3 +1,63 @@
|
|
| 1 |
---
|
|
|
|
| 2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
pretty_name: DMImage Test Set
|
| 3 |
license: apache-2.0
|
| 4 |
+
task_categories:
|
| 5 |
+
- image-classification
|
| 6 |
+
tags:
|
| 7 |
+
- ai-generated-image-detection
|
| 8 |
+
- synthetic-image-detection
|
| 9 |
+
- deepfake-detection
|
| 10 |
+
- benchmark
|
| 11 |
---
|
| 12 |
+
|
| 13 |
+
# DMImage Test Set
|
| 14 |
+
|
| 15 |
+
A constructed test set for **AI-generated image detection**, containing **21,000 images**:
|
| 16 |
+
**5,000 real** photographs and **16,000 synthetic** images produced by 30 generator
|
| 17 |
+
configurations spanning GANs, diffusion models, and autoencoder-based generators.
|
| 18 |
+
|
| 19 |
+
Images are grouped as `TestSet/<subset>/`. Subsets whose name starts with `real_`
|
| 20 |
+
contain real photographs; all other subsets contain images generated by the model
|
| 21 |
+
indicated by the folder name.
|
| 22 |
+
|
| 23 |
+
## Composition
|
| 24 |
+
|
| 25 |
+
| Family | Subsets | Images |
|
| 26 |
+
|---|---|---:|
|
| 27 |
+
| Real | `real_imagenet_val`, `real_coco_valid`, `real_ucid` | 5,000 |
|
| 28 |
+
| GAN | `progan_lsun`, `biggan_256/512`, `stylegan2_*`, `stylegan3_*`, `eg3d` | 5,000 |
|
| 29 |
+
| Diffusion | `stable_diffusion_256`, `latent-diffusion_*`, `guided-diffusion_*`, `glide_text2img_valid`, `dalle-mini_valid`, `dalle_2` | 9,000 |
|
| 30 |
+
| Autoencoder | `taming-transformers_*` | 2,000 |
|
| 31 |
+
|
| 32 |
+
- Real: ImageNet val (2,000, `.JPEG`), COCO val (2,000, `.jpg`), UCID (1,000, `.tif`)
|
| 33 |
+
- Synthetic: `.png`, resolution indicated by the folder name (256 / 512 / 1024)
|
| 34 |
+
|
| 35 |
+
## operations.csv
|
| 36 |
+
|
| 37 |
+
`TestSet/operations.csv` records, for every image, the randomized crop and JPEG
|
| 38 |
+
quality parameters used in the robustness evaluation protocol:
|
| 39 |
+
|
| 40 |
+
| Column | Meaning |
|
| 41 |
+
|---|---|
|
| 42 |
+
| `src` | image path relative to `TestSet/` |
|
| 43 |
+
| `cropsize`, `x1`, `y1` | crop size and top-left offset |
|
| 44 |
+
| `qf` | JPEG quality factor |
|
| 45 |
+
| `typ` | subset name |
|
| 46 |
+
| `label` | `REAL` for real images, `FALSE` for generated images |
|
| 47 |
+
|
| 48 |
+
## Usage
|
| 49 |
+
|
| 50 |
+
```python
|
| 51 |
+
from pathlib import Path
|
| 52 |
+
from PIL import Image
|
| 53 |
+
|
| 54 |
+
root = Path("TestSet")
|
| 55 |
+
for path in root.glob("*/*"):
|
| 56 |
+
is_real = path.parent.name.startswith("real_")
|
| 57 |
+
image = Image.open(path)
|
| 58 |
+
```
|
| 59 |
+
|
| 60 |
+
This is an evaluation-only test set; it is not intended for training.
|
| 61 |
+
|
| 62 |
+
Please respect the licenses of the original image sources (ImageNet, COCO, UCID)
|
| 63 |
+
and of the generative models used to synthesize the images.
|