Add SIDBench pretrained detector weights (34 files) and model card

README.md

@@ -0,0 +1,211 @@
+---
+license: other
+license_name: mixed-upstream-licenses
+license_link: https://github.com/mever-team/sidbench/blob/main/LICENSE
+pipeline_tag: image-classification
+tags:
+  - synthetic-image-detection
+  - ai-generated-image-detection
+  - deepfake-detection
+  - image-forensics
+  - sidbench
+  - benchmark
+  - pytorch
+---
+
+# SIDBench — Pretrained Detector Weights
+
+This repository hosts the pretrained weights used by
+[**SIDBench**](https://github.com/mever-team/sidbench), a framework for benchmarking
+state-of-the-art synthetic image detection methods.
+
+It replaces the Google Drive archive that the SIDBench README used to link to, which is
+no longer reachable. The directory layout here mirrors exactly what the framework expects,
+so the weights can be dropped into a SIDBench checkout with a single command.
+
+**34 files, 8.7 GB total.**
+
+## Download
+
+From the root of a [SIDBench](https://github.com/mever-team/sidbench) checkout:
+
+```bash
+hf download dkarageo/sidbench --local-dir weights
+```
+
+That produces `./weights/<method>/<checkpoint>`, matching the default `--ckpt` paths
+baked into `options/options.py` and `models/models.py` — no further configuration needed.
+
+The CLI ships with `huggingface_hub`:
+
+```bash
+pip install -U "huggingface_hub[cli]"
+```
+
+### Downloading only what you need
+
+The full set is 8.7 GB, but most methods need only a few hundred MB. Fetch a subset with
+`--include` (repeat the flag per pattern):
+
+```bash
+# UnivFD only (4 KB) -- the framework's default model
+hf download dkarageo/sidbench --local-dir weights --include "univfd/*"
+
+# Rine + PatchCraft + NPR (~106 MB)
+hf download dkarageo/sidbench --local-dir weights \
+    --include "rine/*" --include "rptc/*" --include "npr/*"
+```
+
+Note that `Dire` additionally needs `preprocessing/lsun_bedroom.pt` (2.1 GB), `LGrad`
+needs `preprocessing/karras2019stylegan-bedrooms-256x256_discriminator.pth` (88 MB), and
+`DeFake` needs both files under `defake/` plus `defake/model_base_capfilt_large.pth`
+(2.0 GB). See the table below.
+
+## Contents
+
+Sizes are binary (MiB/GiB). The **Consumed by** column gives the SIDBench flag that
+points at each file.
+
+### Detector checkpoints
+
+| File | Size | Consumed by | Trained on | md5 |
+|---|---|---|---|---|
+| `cnndetect/blur_jpg_prob0.1.pth` | 269.4 MB | `--modelName=CNNDetect --ckpt` | proGAN, augmented (recompressed) with 10% probability | `109a7a7406f52a3f658dcd982b58f333` |
+| `cnndetect/blur_jpg_prob0.5.pth` | 269.4 MB | `--modelName=CNNDetect --ckpt` | proGAN, augmented (recompressed) with 50% probability | `0c0bd6f572eaec0e8ea8ed8b033969dd` |
+| `dimd/corvi22_latent_model.pth` | 269.5 MB | `--modelName=DIMD --ckpt` | Latent Diffusion | `56fc46cd42550fe1b4ed819be26e2bfd` |
+| `dimd/corvi22_progan_model.pth` | 269.5 MB | `--modelName=DIMD --ckpt` | proGAN | `a1163e6633acc6f7aac81dbbafa6d3b0` |
+| `dimd/gandetection_resnet50nodown_progan.pth` | 269.4 MB | `--modelName=DIMD --ckpt` | proGAN | `f56d78a7092453e3b3b55bd285c38027` |
+| `dimd/gandetection_resnet50nodown_stylegan2.pth` | 269.5 MB | `--modelName=DIMD --ckpt` | styleGAN2 | `a18f559b83ecac595e1588eea29999ad` |
+| `dire/lsun_adm.pth` | 269.4 MB | `--modelName=Dire --ckpt` | ADM (diffusion) | `02bfd4b29c97e15e82777c474dc3c81f` |
+| `dire/lsun_iddpm.pth` | 269.4 MB | `--modelName=Dire --ckpt` | IDDPM | `41c57afccc6a13bddd4b1c99366577e4` |
+| `dire/lsun_pndm.pth` | 269.4 MB | `--modelName=Dire --ckpt` | PNDM | `7841ebc388f5690c400b9c48aaf71e32` |
+| `dire/lsun_stylegan.pth` | 269.4 MB | `--modelName=Dire --ckpt` | styleGAN | `eaf1839b23a0b3ab308ad7e8d4e4956e` |
+| `freqdetect/DCTAnalysis.pth` | 90.0 MB | `--modelName=FreqDetect --ckpt` | — | `02c3d38cad027db02baf3564722ae6f4` |
+| `fusing/PSM.pth` | 283.8 MB | `--modelName=Fusing --ckpt` | — | `64a67251abf0c501cbf53b1436b5ec7e` |
+| `gramnet/Gram.pth` | 44.9 MB | `--modelName=GramNet --ckpt` | — | `71e8d0aeb1030c959506cf3e45736a67` |
+| `lgrad/LGrad.pth` | 269.5 MB | `--modelName=LGrad --ckpt` | proGAN | `69fdb9f9f8ad10ad33c183a56151e01c` |
+| `lgrad/LGrad-1class-Trainon-Progan_horse.pth` | 90.0 MB | `--modelName=LGrad --ckpt` | proGAN, one class | `16f34fa71ac44b574c8dea1f0bb3179f` |
+| `lgrad/LGrad-2class-Trainon-Progan_chair_horse.pth` | 90.0 MB | `--modelName=LGrad --ckpt` | proGAN, two classes | `918f9a66bc0141c73d4e21f389532cdc` |
+| `lgrad/LGrad-4class-Trainon-Progan_car_cat_chair_horse.pth` | 90.0 MB | `--modelName=LGrad --ckpt` | proGAN, four classes | `e40903550e6530a35d6b2f74b8d076f4` |
+| `npr/NPR.pth` | 16.6 MB | `--modelName=NPR --ckpt` | — | `35d0f34154358af6b38157154b443f53` |
+| `rine/model_1class_trainable.pth` | 40.1 MB | `--modelName=Rine --ckpt` | proGAN, one class | `ef625cfaf25ee6c4a77c70064fda3443` |
+| `rine/model_2class_trainable.pth` | 1.1 MB | `--modelName=Rine --ckpt` | proGAN, two classes | `2959eb954e4afa1e2a6222a8ad6f7bf3` |
+| `rine/model_4class_trainable.pth` | 24.1 MB | `--modelName=Rine --ckpt` | proGAN, four classes | `8931ede3fa2f6f6e98df0dd1563fa946` |
+| `rine/model_ldm_trainable.pth` | 40.1 MB | `--modelName=Rine --ckpt` | Latent Diffusion, one class | `ce0e5a4ea018b49511ec1f9972b8487a` |
+| `rptc/RPTC.pth` | 500.7 KB | `--modelName=PatchCraft --ckpt` | proGAN | `271ec9c97551ab2ce19e1d8bb6545059` |
+| `univfd/fc_weights.pth` | 4.0 KB | `--modelName=UnivFD --ckpt` | proGAN | `392b2d8b637e932a2534ded56d9185bd` |
+| `defake/clip_linear.pth` | 2.5 MB | `--modelName=DeFake --ckpt` | hybrid image+text detector, diffusion images | `bba621d9877a5a795bdb1a0670d0ae5e` |
+
+`Rine` infers its architecture from the checkpoint filename (`model_<ncls>_trainable.pth`),
+so do not rename those four files.
+
+### Auxiliary networks
+
+These are not detectors; they are feature extractors and generative models that some
+methods run as a preprocessing step.
+
+| File | Size | Consumed by | md5 |
+|---|---|---|---|
+| `freqdetect/dct_mean.zip` | 1.1 MB | `--dctMean` (FreqDetect) | `19daa38673b9e2e7c2125e3c81080738` |
+| `freqdetect/dct_var.zip` | 1.1 MB | `--dctVar` (FreqDetect) | `dc31d4f90068d15a5d79cce470955941` |
+| `preprocessing/karras2019stylegan-bedrooms-256x256_discriminator.pth` | 88.0 MB | `--LGradGenerativeModelPath` (LGrad) | `12b5b30f3386cb09692757224e124795` |
+| `preprocessing/lsun_bedroom.pt` | 2.1 GB | `--DireGenerativeModelPath` (Dire) | `34d5da60938c66eca1f327d17f54acfa` |
+| `defake/finetune_clip.pt` | 337.3 MB | `--defakeClipEncoderPath` (DeFake) | `3853db6a3282e60b08d5559a3cef2e2d` |
+| `defake/model_base_capfilt_large.pth` | 2.0 GB | `--defakeBlipPath` (DeFake) | `dd40ed17486a858be6b2e085caba57a8` |
+
+### Not used by the current SIDBench release
+
+Kept for completeness and reproducibility — no code path in SIDBench loads these.
+
+| File | Size | What it is | md5 |
+|---|---|---|---|
+| `preprocessing/karras2019stylegan-bedrooms-256x256.pkl` | 288.5 MB | The original TensorFlow StyleGAN pickle from which the LGrad discriminator above was converted. Unpickling it requires NVIDIA's `dnnlib`; SIDBench only ever loads the converted `_discriminator.pth`. | `56f6480c73c941c59bd4a54653b09ffc` |
+| `lnp/LNP.pth` | 269.5 MB | A ResNet-50 classifier for the LNP method (Liu et al., ECCV 2022). LNP is not in SIDBench's `VALID_MODELS`. | `2fb12ffbf7b685b4751f5fc9c7989e3e` |
+| `preprocessing/sidd_rgb.pth` | 32.6 MB | The SIDD RGB denoising network that LNP uses to extract its learned noise pattern. | `a3e4695fbb32357d2423fd13a4258ee7` |
+
+## Usage notes
+
+A few checkpoints need more than a `--ckpt` path.
+
+**FreqDetect** needs the DCT normalisation statistics. The `--dctMean` / `--dctVar`
+defaults in `options/options.py` omit the `.zip` extension that the distributed files
+carry, so pass both flags explicitly:
+
+```bash
+python test.py --dataPath <images> --modelName=FreqDetect \
+    --ckpt=./weights/freqdetect/DCTAnalysis.pth \
+    --dctMean=./weights/freqdetect/dct_mean.zip \
+    --dctVar=./weights/freqdetect/dct_var.zip
+```
+
+**DeFake** requires the OpenAI CLIP package to be importable as `clip`.
+`defake/finetune_clip.pt` is a pickled `clip.model.CLIP` *object* rather than a
+state dict, so `torch.load` fails with `ModuleNotFoundError: No module named 'clip'`
+without it:
+
+```bash
+pip install git+https://github.com/openai/CLIP.git
+```
+
+**UnivFD** and **Rine** are CLIP-based and download the OpenAI `ViT-L/14` backbone
+(~1.7 GB) to `~/.cache/clip` on first use. That backbone is *not* included here, since
+it is fetched automatically and is not SIDBench-specific. Both also require
+`--resizeSize=224`, as CLIP expects 224×224 inputs.
+
+## Provenance
+
+SIDBench integrates published detectors; the checkpoints below originate with their
+respective authors and are redistributed here so the benchmark stays reproducible.
+
+| Files | Method | Paper | Original code |
+|---|---|---|---|
+| `cnndetect/` | CNNDetect | CNN-generated images are surprisingly easy to spot...for now | [peterwang512/CNNDetection](https://github.com/peterwang512/CNNDetection) |
+| `dimd/` | DIMD | On the detection of synthetic images generated by diffusion models | [grip-unina/DMimageDetection](https://github.com/grip-unina/DMimageDetection) |
+| `freqdetect/` | FreqDetect | Leveraging Frequency Analysis for Deep Fake Image Recognition | [RUB-SysSec/GANDCTAnalysis](https://github.com/RUB-SysSec/GANDCTAnalysis) |
+| `fusing/` | Fusing | Fusing global and local features for generalized AI-synthesized image detection | [littlejuyan/FusingGlobalandLocal](https://github.com/littlejuyan/FusingGlobalandLocal) |
+| `gramnet/` | GramNet | Global Texture Enhancement for Fake Face Detection In the Wild | [liuzhengzhe/Global\_Texture\_Enhancement...](https://github.com/liuzhengzhe/Global_Texture_Enhancement_for_Fake_Face_Detection_in_the-Wild) |
+| `lgrad/`, `preprocessing/karras*` | LGrad | Learning on Gradients: Generalized Artifacts Representation for GAN-Generated Images Detection | [chuangchuangtan/LGrad](https://github.com/chuangchuangtan/LGrad) |
+| `dire/`, `preprocessing/lsun_bedroom.pt` | Dire | DIRE for Diffusion-Generated Image Detection | [ZhendongWang6/DIRE](https://github.com/ZhendongWang6/DIRE) |
+| `univfd/` | UnivFD | Towards Universal Fake Image Detectors that Generalize Across Generative Models | [Yuheng-Li/UniversalFakeDetect](https://github.com/Yuheng-Li/UniversalFakeDetect) |
+| `npr/` | NPR | Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection | [chuangchuangtan/NPR-DeepfakeDetection](https://github.com/chuangchuangtan/NPR-DeepfakeDetection) |
+| `rptc/` | PatchCraft | PatchCraft: Exploring Texture Patch for Efficient AI-generated Image Detection | [project page](https://fdmas.github.io/AIGCDetect/) |
+| `defake/` | DeFake | DE-FAKE: Detection and Attribution of Fake Images Generated by Text-to-Image Generation Models | [zeyangsha/De-Fake](https://github.com/zeyangsha/De-Fake) |
+| `rine/` | Rine | Leveraging Representations from Intermediate Encoder-blocks for Synthetic Image Detection | [mever-team/rine](https://github.com/mever-team/rine) |
+| `lnp/`, `preprocessing/sidd_rgb.pth` | LNP *(not integrated)* | Detecting Generated Images by Real Images | [Tangsenghenshou/Detecting-Generated-Images-by-Real-Images](https://github.com/Tangsenghenshou/Detecting-Generated-Images-by-Real-Images) |
+
+## Integrity
+
+Every file published here was verified with PyTorch 2.3.1 before upload: each detector
+checkpoint was loaded into the SIDBench model class that consumes it via the framework's
+own `load_weights()` code path, and the auxiliary networks
+(`karras2019...discriminator.pth`, `lsun_bedroom.pt`, `model_base_capfilt_large.pth`)
+were loaded with `strict=True`. The md5 sums above let you confirm a download arrived
+intact — a truncated checkpoint typically surfaces as
+`PytorchStreamReader failed reading zip archive: failed finding central directory`.
+
+## License
+
+The SIDBench framework is released under the
+[Apache 2.0 licence](https://github.com/mever-team/sidbench/blob/main/LICENSE). The
+weights themselves were produced by the authors of the individual methods and remain
+subject to the licence terms of their originating projects, linked in the Provenance
+table above. Consult those projects before redistributing or using these checkpoints
+commercially.
+
+## Citation
+
+If you use SIDBench in your research, please cite:
+
+```bibtex
+@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}
+}
+```
+
+Please also cite the original paper of any individual detector you use.

cnndetect/blur_jpg_prob0.1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a73295ac66f9cb74d558ce3ade46f75e2f2997ed05eeed0f4b774623372058ea
+size 282442597

cnndetect/blur_jpg_prob0.5.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3533f3e58a3b82215f21197f2faf4e99b911ec42a4a6e2d7823e04d5f3fd245a
+size 282442597

defake/clip_linear.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:251e5075bb9e7a3df08b0b857686a6cab4bb052ea1b3642b3b757952356bac8c
+size 2629296

defake/finetune_clip.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:607586efd97e17a28700f6f046711b5ed1d5f56bf70270563a26c10d1de695e4
+size 353713101

defake/model_base_capfilt_large.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f5187458d4d47bb87876faf3038d5947eff17475edf52cf47b62e84da0b235f
+size 2115792294

dimd/corvi22_latent_model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0168451e84a2a35d0b141ad4f2e361a74efdd06eb8d11434087f8a0bb6e3abad
+size 282549121

dimd/corvi22_progan_model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25ac8e1614fc3e6ab4a3325948b758e4be6273e713ada71f53d46d9b64070f81
+size 282549121

dimd/gandetection_resnet50nodown_progan.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c9b97ff7bb51d49e23f72e44437bdacce5219afc5a09bbd0619d79ece2d8461
+size 282442597

dimd/gandetection_resnet50nodown_stylegan2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65467594eeb53945417c909390a3d872d55b6dbd819aa12cf01e4ced9c4d5a08
+size 282549121

dire/lsun_adm.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e61a33a6066f23771f75aa717da8e6d7c7ecea6daa3110108368990f55e3e523
+size 282534013

dire/lsun_iddpm.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2869ed6fd083ee6a3ddbcc8ed2b05ecb530a9cb2c050eaefdc976df4fb27752b
+size 282534013

dire/lsun_pndm.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:233836814cf14c2ebfc28b4b2d4f1feec20c66d91f87a33a8e28b7660aca2ea6
+size 282534013

dire/lsun_stylegan.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95efd76f6e0e6cac2f067d2204305c5cac2bc9471a40bcca6015981f2b98f24c
+size 282534013

freqdetect/DCTAnalysis.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ecb4dfa76139a38cfe1c013f433779d3dc61a0ce9a8f262bce2bc90eb0c63f08
+size 94359115

freqdetect/dct_mean.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09740fb73073ff9a8a15f25b9ab7dede6c40c106bbeac40715d8500b14901593
+size 1204974

freqdetect/dct_var.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8213d94ca9274aad678b1df62057b319874b96803d0768a63d73cf9ee233e661
+size 1204971

fusing/PSM.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72274e1d02e6495b0086af5093e86d1dfae340c14da95e8f01d88553f22e54d7
+size 297536830

gramnet/Gram.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc49483b6c988703d94970b3066fdad819d6da5c192fd4c8a601e7dd3470cf69
+size 47057201

lgrad/LGrad-1class-Trainon-Progan_horse.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17c8181b79dec9d0b8648779c30ccbd9838fd6348835f3ebd9a57bc9d9e19629
+size 94364719

lgrad/LGrad-2class-Trainon-Progan_chair_horse.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ec97ce9b4db90ce4b66a88f4d2c93fd6329877d2544c22af4e50d711ae490c4
+size 94364719

lgrad/LGrad-4class-Trainon-Progan_car_cat_chair_horse.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbd4f68698cbe52449ad53eba55f31bba7a9eb20cafeedc37ffe59eecbb13b7d
+size 94364719

lgrad/LGrad.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93daff46f946f7273aabc5ea9b01908c8a548a23365dd3c3c9433b31d104e781
+size 282581704

lnp/LNP.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17cde2d69689361706ad7dd4612959306bd226b201d41790648c3ed576a6267c
+size 282581704

npr/NPR.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3939297e9399e0b992f87211610769d87d899de50d56da0204d6cbda2d483a53
+size 17393733

preprocessing/karras2019stylegan-bedrooms-256x256.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c49c6b716e5b582142884cacb2a93e8e429113b29b55efee585417e94d81ff23
+size 302550530

preprocessing/karras2019stylegan-bedrooms-256x256_discriminator.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad523db88e928bd33dac305560050391f042b0c798c50184770c641e2952fce1
+size 92300155

preprocessing/lsun_bedroom.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9faf136dc2375dcdb392b35cee9ca9dca1fd5257b2f3358613136395ec39231
+size 2211383297

preprocessing/sidd_rgb.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:efff3a8b39c6d65c9d4bb49d95f5fe99126551b0c22e541d7f103e60b307ba51
+size 34228737

rine/model_1class_trainable.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:427621294b40f37f774ef53eff96d205b93c709790ae24f70b4f5687842c5875
+size 42094118

rine/model_2class_trainable.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:680d301b6d191da85b9dabcebf42cd3df20c5c669fa8320c498f835d02bf5ccb
+size 1139686

rine/model_4class_trainable.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6535ff6ecfa88d33c081e561c2cb2dcb594a1006b050eddb72a72ce005ed2ca1
+size 25298182

rine/model_ldm_trainable.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13986ffc86e1bd9d57fd640f891358c4f14c0f8aea8e590e52c598541f69fdc1
+size 42093973

rptc/RPTC.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1c9612cc47215d7c4cbdc5ddff52a4312519d697fd0830ce478f63f58e733fd
+size 512674

univfd/fc_weights.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:477100745713bcc957beb2b40859536859b6483fd6301b3b9293151b194c7847
+size 4083