object2021/aigi-hf-shortcut-data
收藏Hugging Face2026-03-18 更新2026-03-29 收录
下载链接：
https://hf-mirror.com/datasets/object2021/aigi-hf-shortcut-data
下载链接
链接失效反馈官方服务：
资源简介：
---
license: mit
task_categories:
- image-classification
tags:
- aigi-detection
- deepfake
- ai-generated-image
- fingerprint
pretty_name: AIGI HF Shortcut Experiment Data
size_categories:
- 10K<n<100K
---
# AIGI HF Shortcut — Manipulated Image Dataset
Manipulated images for testing whether MLLM-based AI-generated image (AIGI) detectors rely on high-frequency artifacts or semantic reasoning.
## Source
Based on **GenImage** test set (6 generators: ADM, BigGAN, Midjourney, SDv4, SDv5, Wukong).
Original dataset: [jzousz/GenImage](https://huggingface.co/datasets/jzousz/GenImage)
## Conditions
| Condition | Formula | # Images per generator |
|-----------|---------|----------------------|
| **condR** | `real_img` (untouched) | 100 |
| **condA** | `fake_img` (original) | 500 |
| **condB** | `fake_img - α × fingerprint_g` | 500 |
| **condC** | `real_img + α × fingerprint_g` | 500 |
| **condD** | `real_img - α × F_real` | 500 |
- α per generator: ADM/SDV4=0.5, BigGAN/Midjourney/SDV5=0.3, Wukong=0.2
- All images resized to 256×256
- SSIM > 0.95 maintained for all manipulations
## Structure
```
manipulated/
├── condR/{ADM,biggan,midjourney,SDV4,SDV5,Wukong}/
├── condA/{ADM,biggan,midjourney,SDV4,SDV5,Wukong}/
├── condB/{ADM,biggan,midjourney,SDV4,SDV5,Wukong}/
├── condC/{ADM,biggan,midjourney,SDV4,SDV5,Wukong}/
└── condD/{ADM,biggan,midjourney,SDV4,SDV5,Wukong}/
```
## Related
- Code & results: [GWB21/aigi-hf-shortcut](https://github.com/GWB21/aigi-hf-shortcut)
- ManiFPT (Song et al., CVPR 2024): fingerprint extraction method
提供机构：
object2021
