| --- |
| license: mit |
| library_name: pytorch |
| pipeline_tag: image-classification |
| tags: |
| - ai-generated-image-detection |
| - deepfake-detection |
| - dinov3 |
| - cvpr2026 |
| --- |
| |
| # SimLBR: Learning to Detect Fake Images by Learning to Detect Real Images (CVPR 2026) |
|
|
| [](https://arxiv.org/abs/2602.20412) |
| [](https://github.com/mvrl/SimLBR) |
|
|
| Pretrained AI-generated image detectors from the CVPR 2026 paper **SimLBR**. A frozen DINOv3 ViT-L/16 backbone (LVD-1689M weights) produces one CLS embedding per image, and a small MLP head, trained with Latent Blending Regularization (LBR), predicts whether the image is real or fake. |
|
|
| Each file in this repo holds **only the MLP head** (~2.6 MB). The DINOv3 backbone is not redistributed here; it is downloaded from Meta under the [DINOv3 license](https://ai.meta.com/resources/models-and-libraries/dinov3-downloads/). |
|
|
| ## Models |
|
|
| | Variant | Files | Training data | Benchmark | Avg. Acc | Avg. AP | |
| |---|---|---|---|---|---| |
| | `aigc` | `aigc/` | ProGAN | AIGCDetectionBenchMark (16 subsets) | 88.4 | 96.1 | |
| | `genimage` | `genimage/` | GenImage SD v1.4 | GenImage (8 subsets) | 91.1 | 97.1 | |
|
|
| Note: the GenImage numbers above differ from the paper. The paper's GenImage ADM entry was evaluated on real images only, because a data-loading bug skipped ADM's fake images (`.PNG` extension). The numbers here include all ADM images and are what the released code reproduces. |
|
|
| Head: `Linear(1024→512) → ReLU → Dropout → Linear(512→256) → ReLU → Dropout → Linear(256→1)`, trained with LBR α ~ U(0.5, 0.8). Output is a logit; `sigmoid(logit) > 0.5` means fake. |
|
|
| ## Usage |
|
|
| Install the [SimLBR code](https://github.com/mvrl/SimLBR) and set `DINO_V3_KEY` to your DINOv3 ViT-L/16 download URL from Meta: |
|
|
| ```bash |
| git clone https://github.com/mvrl/SimLBR && cd SimLBR |
| export DINO_V3_KEY=YOUR_DINOV3_VITL16_URL |
| ``` |
|
|
| Score images: |
|
|
| ```bash |
| python -m simlbr.predict --pretrained aigc image1.png image2.jpg |
| ``` |
|
|
| Load in Python: |
|
|
| ```python |
| from simlbr import SimLBR |
| |
| model = SimLBR.from_pretrained("aigc") # or "genimage" |
| ``` |
|
|
| Reproduce benchmark numbers: |
|
|
| ```bash |
| python -m simlbr.evaluate --dataset_name aigc --data_dir /path/to/AIGCDetectionBenchMark --pretrained aigc |
| ``` |
|
|
| **Preprocessing matters:** images must be resized to 256×256 and normalized with `mean=(0.430, 0.411, 0.296)`, `std=(0.213, 0.156, 0.143)`, as in `simlbr.utils.make_dinov3_transform()`. The heads were trained with these exact statistics. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{dhakal2026simlbr, |
| title={SimLBR: Learning to Detect Fake Images by Learning to Detect Real Images}, |
| author={Dhakal, Aayush and Khanal, Subash and Sastry, Srikumar and Arndt, Jacob and Ambrozio Dias, Philipe and Lunga, Dalton and Jacobs, Nathan}, |
| booktitle={Computer Vision and Pattern Recognition}, |
| year={2026}, |
| organization={IEEE/CVF} |
| } |
| ``` |
|
|
