Effort-AIGI-Detection/UniversalFakeDetect_Benchmark/README_FOR_YOU.txt

Quick start for your REAL vs FAKE dataset
=========================================

Your dataset structure (what this patch expects):
  <DATASET_ROOT>/
    train/
      real/
      fake/
    test/
      real/
      fake/

What was changed:
  1) data/datasets.py
     - Validation now reads from "test" (instead of "test/progan").
     - The loader looks for folder names containing "real" and "fake" (instead of "0_real" and "1_fake").

  2) models/__init__.py
     - Pretrained model names now use Hugging Face IDs so they auto-download (no internal file paths needed).

Environment (example):
  pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
  pip install transformers loralib tensorboardX scikit-image opencv-python scikit-learn pillow tqdm

Train example (single GPU):
  cd UniversalFakeDetect_Benchmark
  python train.py \
    --name clip_vitl14_svd \
    --wang2020_data_path <DATASET_ROOT> \
    --data_mode wang2020 \
    --arch "CLIP:ViT-L/14_svd" \
    --batch_size 32 \
    --loadSize 256 \
    --cropSize 224 \
    --lr 2e-4 \
    --use_svd

Notes:
  - Checkpoints and logs will go under: checkpoints/<run name>/
  - For CPU-only training (slow), add: --gpu_ids -1
  - You can try other backbones by setting --arch to one of:
      "CLIP:ViT-B/16_svd", "CLIP:ViT-B/32_svd", "CLIP:ViT-L/14_svd",
      "SigLIP:ViT-L/16_256_svd", "BEiTv2:ViT-L/16_svd"
  - To disable extra augmentations, do NOT pass --data_aug (or remove it if present).

Validation:
  The training script automatically builds a validation loader with data_label='val'
  and (with this patch) points it to <DATASET_ROOT>/test.

In case you prefer ZERO code changes:
  - Keep the original repo, but rename your folders to:
      train/0_real, train/1_fake, test/progan/0_real, test/progan/1_fake
    Then you can run the unpatched code with --data_mode wang2020.