SIDBench: A Python Framework for Reliably Assessing Synthetic Image Detection Methods
Paper • 2404.18552 • Published
SIDBench — Pretrained Detector Weights
Pretrained weights for SIDBench.
31 checkpoints, 8.74 GB.
Download
From the root of a SIDBench checkout:
pip install -U "huggingface_hub[cli]"
hf download dkarageo/sidbench --local-dir weights
This yields ./weights/<method>/<checkpoint>, matching the default --ckpt paths in
options/options.py and models/models.py.
If you would like to run only a subset of the models, download just what they need, with
one --include per pattern:
# UnivFD only (4.1 kB)
hf download dkarageo/sidbench --local-dir weights --include "univfd/*"
# Rine + PatchCraft + NPR (~129 MB)
hf download dkarageo/sidbench --local-dir weights \
--include "rine/*" --include "rptc/*" --include "npr/*"
Dire, LGrad and DeFake each need an additional file from preprocessing/ or
defake/, listed under Auxiliary networks below.
Contents
Detector checkpoints
Each is passed via --ckpt, together with the matching --modelName.
| File | Size (MB = 10⁶ B) | --modelName |
Trained on | md5 |
|---|---|---|---|---|
cnndetect/blur_jpg_prob0.1.pth |
282.4 MB | CNNDetect |
proGAN, augmented (recompressed) with 10% probability | 109a7a7406f52a3f658dcd982b58f333 |
cnndetect/blur_jpg_prob0.5.pth |
282.4 MB | CNNDetect |
proGAN, augmented (recompressed) with 50% probability | 0c0bd6f572eaec0e8ea8ed8b033969dd |
dimd/corvi22_latent_model.pth |
282.5 MB | DIMD |
Latent Diffusion | 56fc46cd42550fe1b4ed819be26e2bfd |
dimd/corvi22_progan_model.pth |
282.5 MB | DIMD |
proGAN | a1163e6633acc6f7aac81dbbafa6d3b0 |
dimd/gandetection_resnet50nodown_progan.pth |
282.4 MB | DIMD |
proGAN | f56d78a7092453e3b3b55bd285c38027 |
dimd/gandetection_resnet50nodown_stylegan2.pth |
282.5 MB | DIMD |
styleGAN2 | a18f559b83ecac595e1588eea29999ad |
dire/lsun_adm.pth |
282.5 MB | Dire |
ADM (diffusion) | 02bfd4b29c97e15e82777c474dc3c81f |
dire/lsun_iddpm.pth |
282.5 MB | Dire |
IDDPM | 41c57afccc6a13bddd4b1c99366577e4 |
dire/lsun_pndm.pth |
282.5 MB | Dire |
PNDM | 7841ebc388f5690c400b9c48aaf71e32 |
dire/lsun_stylegan.pth |
282.5 MB | Dire |
styleGAN | eaf1839b23a0b3ab308ad7e8d4e4956e |
freqdetect/DCTAnalysis.pth |
94.4 MB | FreqDetect |
— | 02c3d38cad027db02baf3564722ae6f4 |
fusing/PSM.pth |
297.5 MB | Fusing |
— | 64a67251abf0c501cbf53b1436b5ec7e |
gramnet/Gram.pth |
47.1 MB | GramNet |
— | 71e8d0aeb1030c959506cf3e45736a67 |
lgrad/LGrad.pth |
282.6 MB | LGrad |
proGAN | 69fdb9f9f8ad10ad33c183a56151e01c |
lgrad/LGrad-1class-Trainon-Progan_horse.pth |
94.4 MB | LGrad |
proGAN, one class | 16f34fa71ac44b574c8dea1f0bb3179f |
lgrad/LGrad-2class-Trainon-Progan_chair_horse.pth |
94.4 MB | LGrad |
proGAN, two classes | 918f9a66bc0141c73d4e21f389532cdc |
lgrad/LGrad-4class-Trainon-Progan_car_cat_chair_horse.pth |
94.4 MB | LGrad |
proGAN, four classes | e40903550e6530a35d6b2f74b8d076f4 |
npr/NPR.pth |
17.4 MB | NPR |
— | 35d0f34154358af6b38157154b443f53 |
rine/model_1class_trainable.pth |
42.1 MB | Rine |
proGAN, one class | ef625cfaf25ee6c4a77c70064fda3443 |
rine/model_2class_trainable.pth |
1.1 MB | Rine |
proGAN, two classes | 2959eb954e4afa1e2a6222a8ad6f7bf3 |
rine/model_4class_trainable.pth |
25.3 MB | Rine |
proGAN, four classes | 8931ede3fa2f6f6e98df0dd1563fa946 |
rine/model_ldm_trainable.pth |
42.1 MB | Rine |
Latent Diffusion, one class | ce0e5a4ea018b49511ec1f9972b8487a |
rptc/RPTC.pth |
512.7 kB | PatchCraft |
proGAN | 271ec9c97551ab2ce19e1d8bb6545059 |
univfd/fc_weights.pth |
4.1 kB | UnivFD |
proGAN | 392b2d8b637e932a2534ded56d9185bd |
defake/clip_linear.pth |
2.6 MB | DeFake |
hybrid image+text detector, diffusion images | bba621d9877a5a795bdb1a0670d0ae5e |
Rine parses model_<ncls>_trainable out of the filename to select its architecture; do
not rename those four files.
Auxiliary networks
| File | Size (MB = 10⁶ B) | Consumed by | md5 |
|---|---|---|---|
freqdetect/dct_mean.zip |
1.2 MB | --dctMean (FreqDetect) |
19daa38673b9e2e7c2125e3c81080738 |
freqdetect/dct_var.zip |
1.2 MB | --dctVar (FreqDetect) |
dc31d4f90068d15a5d79cce470955941 |
preprocessing/karras2019stylegan-bedrooms-256x256_discriminator.pth |
92.3 MB | --LGradGenerativeModelPath (LGrad) |
12b5b30f3386cb09692757224e124795 |
preprocessing/lsun_bedroom.pt |
2.21 GB | --DireGenerativeModelPath (Dire) |
34d5da60938c66eca1f327d17f54acfa |
defake/finetune_clip.pt |
353.7 MB | --defakeClipEncoderPath (DeFake) |
3853db6a3282e60b08d5559a3cef2e2d |
defake/model_base_capfilt_large.pth |
2.12 GB | --defakeBlipPath (DeFake) |
dd40ed17486a858be6b2e085caba57a8 |
Provenance
| Files | Method | Paper | Original code |
|---|---|---|---|
cnndetect/ |
CNNDetect | CNN-generated images are surprisingly easy to spot...for now | peterwang512/CNNDetection |
dimd/ |
DIMD | On the detection of synthetic images generated by diffusion models | grip-unina/DMimageDetection |
freqdetect/ |
FreqDetect | Leveraging Frequency Analysis for Deep Fake Image Recognition | RUB-SysSec/GANDCTAnalysis |
fusing/ |
Fusing | Fusing global and local features for generalized AI-synthesized image detection | littlejuyan/FusingGlobalandLocal |
gramnet/ |
GramNet | Global Texture Enhancement for Fake Face Detection In the Wild | liuzhengzhe/Global_Texture_Enhancement... |
lgrad/, preprocessing/karras* |
LGrad | Learning on Gradients: Generalized Artifacts Representation for GAN-Generated Images Detection | chuangchuangtan/LGrad |
dire/, preprocessing/lsun_bedroom.pt |
Dire | DIRE for Diffusion-Generated Image Detection | ZhendongWang6/DIRE |
univfd/ |
UnivFD | Towards Universal Fake Image Detectors that Generalize Across Generative Models | Yuheng-Li/UniversalFakeDetect |
npr/ |
NPR | Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection | chuangchuangtan/NPR-DeepfakeDetection |
rptc/ |
PatchCraft | PatchCraft: Exploring Texture Patch for Efficient AI-generated Image Detection | project page |
defake/ |
DeFake | DE-FAKE: Detection and Attribution of Fake Images Generated by Text-to-Image Generation Models | zeyangsha/De-Fake |
rine/ |
Rine | Leveraging Representations from Intermediate Encoder-blocks for Synthetic Image Detection | mever-team/rine |
Integrity
Every checkpoint was loaded into the SIDBench model class that consumes it, with PyTorch 2.3.1, before upload. md5 sums are listed above.
License
SIDBench is Apache 2.0. The weights were produced by the authors of the individual methods and remain subject to the licence terms of the projects listed under Provenance.
Citation
@inproceedings{schinas2024sidbench,
title = {SIDBench: A Python framework for reliably assessing synthetic image detection methods},
author = {Schinas, Manos and Papadopoulos, Symeon},
booktitle = {Proceedings of the 3rd ACM International Workshop on Multimedia AI against Disinformation (MAD '24)},
year = {2024},
doi = {10.1145/3643491.3660277},
eprint = {2404.18552},
archivePrefix = {arXiv}
}
Cite also the original paper of any individual detector used.