heisen111/deepfake_challenge3

library_name: pytorch tags: - deepfake-detection - image-forensics - localization - ddl-x - swin-transformer datasets: - DDL-X metrics: - accuracy - iou

DDL-X Deepfake Detection: Swin-Base Trace-Head Checkpoint

This repository publishes a trained checkpoint for DDL-X Track 3 deepfake detection, localization, and visible forgery trace prediction.

Model Details

• Architecture: Swin-Base visual backbone with classification, localization, and trace-head outputs
• Input size: 384 x 384
• Checkpoint file: swin_base_384_ep100_pretrained_trace_head_best.pt
• Checkpoint size: 1019 MB
• SHA256: 725fd1f9ed1baec3d8c10ea97918981c50a0ccf1fc39fc9e0649c3d14566182e
• Source checkpoint path: /mnt/big_disk/qi.wang/ddl_deepfake/runs/swin_base_384_ep100_pretrained_trace_head/best.pt

Intended Use

The model is intended for research on DDL-X-style image deepfake analysis. It predicts:

• image-level real / fake classification;
• manipulated-region localization masks, exported by the code as challenge-format bounding boxes;
• visible forgery trace keywords/text through the trace-head pipeline.

Usage

Clone or download the lightweight code included in code/, install dependencies, and run inference:

pip install -r code/requirements.txt

python code/infer.py \
  --data-root /path/to/DDL_X \
  --image-dir /path/to/DDL_X/image \
  --checkpoint swin_base_384_ep100_pretrained_trace_head_best.pt \
  --out-dir submissions/json \
  --amp

Verify the checkpoint after download:

sha256sum swin_base_384_ep100_pretrained_trace_head_best.pt

Expected:

725fd1f9ed1baec3d8c10ea97918981c50a0ccf1fc39fc9e0649c3d14566182e

Training Data

The model was trained on the local DDL-X development split generated by code/scripts/split_dataset.py.

Split	Images	Fake Images
Train	93,226	52,297
Validation	10,359	5,811
Total paired samples	103,585	58,108

Limitations

This checkpoint was trained for the DDL-X challenge data distribution and should not be treated as a general-purpose forensic detector without further validation. Localization targets are derived from coarse bounding boxes, so predicted masks and boxes may not precisely follow manipulation boundaries.

Citation

If you use this checkpoint, please cite the DDL-X Track 3 workshop/challenge page and this repository.