| --- |
| license: apache-2.0 |
| pipeline_tag: image-classification |
| library_name: pytorch |
| base_model: openai/clip-vit-large-patch14 |
| tags: |
| - ai-generated-image-detection |
| - deepfake-detection |
| - synthetic-image-detection |
| - image-forensics |
| - clip |
| - vision-language-model |
| - prompt-learning |
| - pytorch-lightning |
| - arxiv:2408.08412 |
| --- |
| |
| # PoundNet |
|
|
| PoundNet checkpoint weights for the paper **"Penny-Wise and Pound-Foolish in AI-Generated Image Detection"**. |
|
|
| PoundNet is a CLIP-based AI-generated image detector built around asymmetric prompt learning for binary real/fake classification and category-aware supervision. The method is designed to reduce the "penny-wise and pound-foolish" behavior of deepfake detectors: strong performance on a narrow training distribution but poor generalization and degraded upstream semantic knowledge. |
|
|
| These weights are released for use with the official PoundNet codebase: |
|
|
| - Code: https://github.com/iamwangyabin/PoundNet |
| - arXiv: https://arxiv.org/abs/2408.08412 |
| - Model weights: https://huggingface.co/nebula/PoundNet |
|
|
| ## Model Details |
|
|
| - **Architecture**: PoundNet |
| - **Backbone**: CLIP ViT-L/14 |
| - **Task**: binary AI-generated image detection / deepfake detection |
| - **Output**: real/fake prediction scores through the official evaluation code |
| - **Training data**: ProGAN split from the ForenSynths-style training setup used by the official PoundNet implementation |
| - **Checkpoint format**: PyTorch Lightning `.ckpt` |
|
|
| The checkpoints in this repository are not standalone `transformers` checkpoints. They should be loaded with the official PoundNet repository and configuration files. |
|
|
| ## Released Checkpoints |
|
|
| | Checkpoint | File | |
| | --- | --- | |
| | `poundnet_ViTL_Progan_20240506_23_30_25` | `poundnet_ViTL_Progan_20240506_23_30_25/last.ckpt` | |
| | `poundnet_ViTL_Progan_20240804_21_16_47` | `poundnet_ViTL_Progan_20240804_21_16_47/last.ckpt` | |
| | `poundnet_ViTL_Progan_20240805_10_31_08` | `poundnet_ViTL_Progan_20240805_10_31_08/last.ckpt` | |
| | `poundnet_ViTL_Progan_20240805_12_09_21` | `poundnet_ViTL_Progan_20240805_12_09_21/last.ckpt` | |
|
|
| ## Installation |
|
|
| Clone the official repository and install dependencies: |
|
|
| ```bash |
| git clone https://github.com/iamwangyabin/PoundNet.git |
| cd PoundNet |
| pip install -r requirements.txt |
| ``` |
|
|
| Install PyTorch separately according to your CUDA environment before installing the remaining dependencies. |
|
|
| ## Download Weights |
|
|
| ```bash |
| mkdir -p weights |
| |
| wget -O ./weights/poundnet_ViTL_Progan_20240506_23_30_25.ckpt \ |
| https://huggingface.co/nebula/PoundNet/resolve/main/poundnet_ViTL_Progan_20240506_23_30_25/last.ckpt |
| |
| wget -O ./weights/poundnet_ViTL_Progan_20240804_21_16_47.ckpt \ |
| https://huggingface.co/nebula/PoundNet/resolve/main/poundnet_ViTL_Progan_20240804_21_16_47/last.ckpt |
| |
| wget -O ./weights/poundnet_ViTL_Progan_20240805_10_31_08.ckpt \ |
| https://huggingface.co/nebula/PoundNet/resolve/main/poundnet_ViTL_Progan_20240805_10_31_08/last.ckpt |
| |
| wget -O ./weights/poundnet_ViTL_Progan_20240805_12_09_21.ckpt \ |
| https://huggingface.co/nebula/PoundNet/resolve/main/poundnet_ViTL_Progan_20240805_12_09_21/last.ckpt |
| ``` |
|
|
| ## Evaluation |
|
|
| PoundNet expects benchmark datasets saved in Hugging Face Arrow format and loaded through `datasets.load_from_disk(...)`. See the official repository for the expected dataset layout and download helper. |
|
|
| Example evaluation command: |
|
|
| ```bash |
| python test.py --cfg cfgs/poundnet.yaml \ |
| datasets.base_path=/path/to/DF-arrow |
| ``` |
|
|
| The default evaluation config uses the ViT-L/14 PoundNet checkpoint and evaluates on multiple AI-generated image detection benchmarks through the official codebase. |
|
|
| ## Intended Uses |
|
|
| PoundNet is intended for academic research on: |
|
|
| - AI-generated image detection |
| - deepfake detection |
| - synthetic image forensics |
| - cross-generator and cross-dataset generalization |
| - prompt-learning adaptation of vision-language models |
|
|
| ## Limitations |
|
|
| These checkpoints are research artifacts and should not be treated as a complete production moderation or forensic system. Performance can vary under distribution shifts such as unseen generators, image editing pipelines, social media compression, resizing, screenshots, adversarial post-processing, or domain-specific content. |
|
|
| The released checkpoints require the official PoundNet code and configuration files. They are not directly loadable through `AutoModel.from_pretrained`. |
|
|
| ## Ethical Considerations |
|
|
| PoundNet is released to support research on synthetic media detection and trustworthy image forensics. Users should validate performance carefully before applying it to real-world moderation, legal, journalistic, or security workflows. Detection results should not be used as the sole evidence for high-stakes decisions. |
|
|
| ## Citation |
|
|
| If you use PoundNet, please cite: |
|
|
| ```bibtex |
| @article{wang2026pennywise, |
| title = {Penny-Wise and Pound-Foolish in AI-Generated Image Detection}, |
| author = {Wang, Yabin and Huang, Zhiwu and Su, Zhou and Prugel-Bennett, Adam and Hong, Xiaopeng}, |
| journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence}, |
| pages = {1--14}, |
| year = {2026}, |
| doi = {10.1109/TPAMI.2026.3664388} |
| } |
| ``` |
|
|
| ## Links |
|
|
| - Code: https://github.com/iamwangyabin/PoundNet |
| - arXiv: https://arxiv.org/abs/2408.08412 |
| - Weights: https://huggingface.co/nebula/PoundNet |
|
|
