HSIC

HSIC Bottleneck for Cross-Generator and Domain-Incremental Synthetic Image Detection
Chin-Chia Yang, Yung-Yu Chuang, Hwann-Tzong Chen, Tyng-Luh Liu

ICLR 2026

[Paper]

Contents

• Setup
• Datasets
• Training

Setup

1. Clone this repository

git clone https://github.com/jamesyoung0623/HSIC.git
cd HSIC

2. Create the conda environment from the exported file

conda env create -f environment.yml
conda activate HSIC

Datasets

This repository trains on pre-extracted CLIP features rather than raw images. The expected workflow is:

1. Register the image roots in dataset_paths_train.py and dataset_paths_test.py.
2. Run extract_features_ddp.py to extract CLIP features for each dataset.
3. Run pack_features.py to concatenate the extracted per-layer features into a single packed_features.npy file for each split.

The 3DGS benchmark data is hosted on Hugging Face: https://huggingface.co/datasets/jamesyoung0623/3DGS_Synthetic_Image_Benchmark

One way to download it is with the Hugging Face CLI:

pip install -U "huggingface_hub[cli]"
hf download jamesyoung0623/3DGS_Synthetic_Image_Benchmark \
  --repo-type dataset \
  --local-dir ./datasets

If the dataset requires authentication, log in first:

hf auth login

After download, organize the extracted folders under datasets/train/... and datasets/test/... so they match the layout below and the paths you register in dataset_paths_train.py and dataset_paths_test.py.

For GAN datasets, please refer to CNNDetection.

For diffusion datasets, please refer to GenImage.

The raw image directories referenced in dataset_paths_train.py and dataset_paths_test.py should follow the wang2020 layout used by the extractor:

datasets/
  train/
    <dataset>/
      0_real/
      1_fake/
  test/
    <dataset>/
      0_real/
      1_fake/

Add each dataset to DATASET_PATHS as a dictionary with real_path, fake_path, data_mode, and key. For datasets already organized with 0_real and 1_fake folders, use data_mode='wang2020'.

Example entry:

dict(
    real_path='/path/to/datasets/train/sdv4',
    fake_path='/path/to/datasets/train/sdv4',
    data_mode='wang2020',
    key='sdv4'
)

To extract CLIP ViT-L/14 features with distributed inference:

PYTHONNOUSERSITE=1 python -m torch.distributed.run --standalone --nproc_per_node=2 \
  extract_features_ddp.py \
  --arch "CLIP:ViT-L/14" \
  --batch_size 128 \
  --flush_every 4096

This writes one file per layer into each dataset directory, for example:

datasets/train/sdv4/0_real/feature_layer0.npy
datasets/train/sdv4/0_real/feature_layer1.npy
...
datasets/train/sdv4/0_real/feature_before_projection.npy

pack_features.py then concatenates the layer files into the final packed feature matrix:

python pack_features.py ./datasets/train/sdv4/0_real
python pack_features.py ./datasets/train/sdv4/1_fake
python pack_features.py ./datasets/test/<dataset>/0_real
python pack_features.py ./datasets/test/<dataset>/1_fake

pack_features.py concatenates the default layer set (layer0 to layer23 plus before_projection) and writes:

datasets/<split>/<dataset>/0_real/packed_features.npy
datasets/<split>/<dataset>/1_fake/packed_features.npy

These packed_features.npy files are the inputs consumed by train_HSIC.py and train_HGR.py.

Training

train_HSIC.py

train_HSIC.py trains the HSIC detector from pre-extracted CLIP features stored as packed_features.npy. Before launching training, make sure the training split contains:

datasets/
  train/
    <dataset>/
      0_real/packed_features.npy
      1_fake/packed_features.npy

The script loads datasets/train/<dataset>/0_real/packed_features.npy and datasets/train/<dataset>/1_fake/packed_features.npy, trains a single model with the built-in dataset-specific HSIC weights, saves checkpoints after each epoch, and then evaluates all saved checkpoints on every available packed test dataset under datasets/test/*/{0_real,1_fake}/packed_features.npy.

Example:

CUDA_VISIBLE_DEVICES=0 python train_HSIC.py \
  --name sdv4 \
  --dataset sdv4

Outputs are written to checkpoints/<name>/ and include:

• opt.txt: parsed command-line options.
• hparams.txt: dataset name and the selected lambda_x / lambda_y.
• model_epoch_*.pth: one checkpoint per epoch.
• train/: TensorBoard logs.
• results_<checkpoint>.txt and summary.csv: post-training evaluation across all available test datasets.

--dataset currently accepts progan and sdv4.

train_HGR.py

train_HGR.py runs the domain-incremental HGR training pipeline with replay selection based on k-center coverage and HSIC centrality. It expects packed features for train, val, and test splits:

datasets/
  train/<dataset>/{0_real,1_fake}/packed_features.npy
  val/<dataset>/{0_real,1_fake}/packed_features.npy
  test/<dataset>/{0_real,1_fake}/packed_features.npy

The script preloads all required datasets, initializes from the base checkpoint, and then evaluates all permutations of the tail-domain order using the hyperparameters defined in train_HGR.py.

The current setup uses base domains sdv4 and progan, tail domains nersemble, NHA, and GAGAvatar, and base checkpoints at:

./checkpoints/sdv4/model_best.pth
./checkpoints/progan/model_best.pth

Example:

CUDA_VISIBLE_DEVICES=0 python train_HGR.py \
  --name HGR \
  --base-select progan \
  --keep_frac 0.01

Outputs are written under checkpoints/<name>__bo_<base>__<sequence>__.../ and include model checkpoints, TensorBoard logs, and log.txt with evaluation results. Per-sequence results are written under seq_logs/.

Citation

If you find our work helpful in your research, please cite it using the following:

@inproceedings{
      yang2026hsic,
      title={{HSIC} Bottleneck for Cross-Generator and Domain-Incremental Synthetic Image Detection},
      author={Chin-Chia Yang and Yung-Yu Chuang and Hwann-Tzong Chen and Tyng-Luh Liu},
      booktitle={The Fourteenth International Conference on Learning Representations},
      year={2026},
      url={https://openreview.net/forum?id=msLnKDvhBx}
}