Datasets:
metadata
license: cc-by-4.0
task_categories:
- image-classification
tags:
- image-forensics
- deepfake-detection
- inpainting-detection
- diffusion-models
- contrastive-learning
size_categories:
- 1M<n<10M
DiffusionPrint Patch Dataset
A dataset of 64x64 image patches for contrastive learning of diffusion-based inpainting forensics. Each patch comes from either a real image or an AI-inpainted region generated by one of three diffusion models.
Columns
| Column | Type | Description |
|---|---|---|
image |
bytes (PNG) |
Lossless 64x64 RGB patch |
master_index |
int64 |
Row index in the original memmap archive |
patch_path |
string |
Original relative path of the patch |
category |
string |
real or generated |
generator_model |
string |
none (real), sd2, sdxl, or flux |
has_positive |
bool |
Whether this patch has a contrastive positive counterpart |
positive_master_indices |
string |
Semicolon-separated master_index values of positive pairs |
Pair Structure
Positive pairs are patches from the same spatial location under different conditions
(e.g. a real patch and its diffusion-reconstructed counterpart).
positive_master_indices stores semicolon-separated integer indices that resolve directly
against the master_index column -- no filesystem paths needed.
Generator Labels
generator_model |
Label int |
|---|---|
none (real) |
0 |
sd2 |
1 |
sdxl |
2 |
flux |
3 |
Stats
- Total patches: 3,688,480
- Shards: 37
- Patch size: 64x64 RGB, lossless PNG