TwinD / README.md
yermandy's picture
TwinD training set: Parquet shards and dataset card
622c8cb verified
|
Raw History Blame Contribute Delete
4.12 kB
---
license: other
license_name: twind-research-only
pretty_name: TwinD training set
task_categories:
- image-classification
tags:
- deepfake-detection
- face-forgery
- faceforensics
size_categories:
- 10K<n<100K
configs:
- config_name: default
data_files:
- split: train
path: data/*/*/*.parquet
extra_gated_prompt: >-
The frames of this dataset are derived from FaceForensics++ and Casual Conversations v2. By requesting access you
confirm that you have accepted the terms of use of both datasets and that you will use this data for non-commercial
research only.
extra_gated_fields:
Name: text
Affiliation: text
I have accepted the FaceForensics++ and Casual Conversations v2 terms of use: checkbox
I will use this data for non-commercial research only: checkbox
---
# TwinD training set
Face crops used to train [TwinD](https://huggingface.co/yermandy/TwinD), from the paper *TwinD: Generalizing
Deepfake Detection from Blending to Entire Face Synthesis*.
- **Reals**: 8 frames of every FaceForensics++ training video (5,739) and 1 random frame of 1,913 Casual
Conversations v2 videos.
- **FF++ fakes**: the Deepfakes, Face2Face, FaceSwap and NeuralTextures frames of the same FF++ frames (22,907), each
cropped with the face box of its real frame, so the two crops match pixel by pixel.
- **Reconstructions**: every real frame passed through an autoencoder, pixel-aligned with it. CogVideoX is the one of
TwinD; KVAE 2D, LTX-Video, HunyuanVideo 1.0 and LTX-2 are the ablations of Tab. 3 of the paper.
Faces are detected with SCRFD-10GF, aligned on their 5 keypoints, the face box enlarged 1.3x, and cropped at about
their original size. The twins of TwinD (partial reconstructions and blob twins) are made from these frames during
training.
## Structure
`data/<dataset>/<method>/train-XXXXX-of-YYYYY.parquet`, shards of at most 500 MB, one row per frame:
| column | content |
| ------------- | ------------------------------------------------------------------------ |
| `image` | the PNG crop (bytes, as stored) |
| `path` | e.g. `FF-x1.3-th0.5-all-aligned/real/001/frame_0000.png` |
| `label` | `real` / `fake` |
| `dataset` | `FF++` or `CCv2` |
| `method` | `real`, `DF`, `F2F`, `FS`, `NT`, `cogvideox`, `kvae2d`, `ltxvideo`, `hunyuanvideo`, `ltx2` |
| `video` | the video of the frame |
| `paired_real` | `path` of the real frame a fake was made from (empty for reals) |
| dataset | method | frames |
| ------- | ---------------------------------------------------------- | -----: |
| FF++ | real | 5,739 |
| FF++ | DF / F2F / FS / NT | 22,907 |
| FF++ | cogvideox / kvae2d / ltxvideo / hunyuanvideo / ltx2, each | 5,739 |
| CCv2 | real | 1,913 |
| CCv2 | cogvideox / kvae2d / ltxvideo / hunyuanvideo / ltx2, each | 1,913 |
68,819 frames, 10.8 GB.
## Usage
Accept the terms above, log in (`hf auth login`), and with the [TwinD code](https://huggingface.co/yermandy/TwinD):
```bash
python prepare_data.py # TwinD subsets, extracted to data/ as PNG
python prepare_data.py --decoders kvae2d ltxvideo hunyuanvideo ltx2 # + the ablation autoencoders
```
or directly:
```python
from datasets import load_dataset
ds = load_dataset("yermandy/TwinD", split="train")
```
## Terms of use
The frames are derived from [FaceForensics++](https://github.com/ondyari/FaceForensics) and
[Casual Conversations v2](https://ai.meta.com/datasets/casual-conversations-v2-dataset/) and are subject to their
terms of use. Use this dataset for non-commercial research only.