💎 SAFIRE

August 5, 2026 · View on GitHub

Welcome to the official repository for the paper "SAFIRE: Segment Any Forged Image Region", accepted at AAAI 2025.

SAFIRE specializes in image forgery localization through two methods: binary localization and multi-source partitioning.

Binary localization identifies the forged regions in an image by generating a heatmap that visualizes the probability of each pixel being manipulated.
Multi-source partitioning divides the image into segments based on their originating sources. This task is proposed for the first time in this paper.

📄 Paper

Authors: Myung-Joon Kwon*, Wonjun Lee*, Seung-Hun Nam, Minji Son, and Changick Kim
Title: SAFIRE: Segment Any Forged Image Region
Conference: Proceedings of the AAAI Conference on Artificial Intelligence, 2025

The paper is available on [arXiv Link].

🎨 Example input / output:

🎁 SafireMS Dataset

The SafireMS Dataset is introduced in our paper and is publicly available on Kaggle for RESEARCH PURPOSES ONLY:

SafireMS-Auto: Automatically generated datasets used for pretraining.
SafireMS-Expert: Manually created datasets designed for evaluating multi-source partitioning performance.

📦️ Training Dataset

We follow the CAT-Net settings for training, which means we used the following datasets:

FantasticReality: The official link is broken. Use: [Link to Download]
CASIA: The official link is broken. Use: [Link to Download]
IMD2020: Official link: [Link to Download]
TampCOCO: Official link: [Link to Download]

Note that we exclude CompRAISE from the original CAT-Net settings.

⚙️ Setup

Clone the repository

git clone https://github.com/mjkwon2021/SAFIRE.git
cd SAFIRE

Download pre-trained weights
Download the weights from [Google Drive Link].
Place the downloaded weights in the root directory of this repository.
Install dependencies
```
conda env create -f environment.yaml
conda activate safire
```
For manual installation, run the commands listed in manual_env_setup.txt.

🚀 Inference

SAFIRE supports two inference types: binary forgery localization and multi-source partitioning.

Prepare Input Images
- Place your input images in the directory: ForensicsEval/inputs.
Output Locations
- Outputs for binary forgery localization will be saved in: ForensicsEval/outputs_binary.
- Outputs for multi-source partitioning will be saved in: ForensicsEval/outputs_multi.

Binary Forgery Localization

Run the following command:

python infer_binary.py --resume="safire.pth"

Multi-Source Partitioning

Using k-means clustering:

python infer_multi.py --resume="safire.pth" --cluster_type="kmeans" --kmeans_cluster_num=3

Using DBSCAN clustering:

python infer_multi.py --resume="safire.pth" --cluster_type="dbscan" --dbscan_eps=0.2 --dbscan_min_samples=1

🧪 Test

To evaluate the model on your test dataset:

Download the test dataset
Obtain the test dataset and place it in a desired location.
Set the dataset path
Update the dataset path in ForensicsEval/project_config.py to point to your downloaded dataset.

Run the evaluation

For binary prediction:

python test_binary.py --resume="safire.pth"

For multi-source partitioning:

python test_multi.py --resume="safire.pth" --cluster_type="kmeans" --kmeans_cluster_num=3

View Results
The evaluation results will be saved as an Excel file.

🔩 Pretrain

We provide support for distributed data parallel (DDP) pretraining of the SAFIRE image encoder. You need SafireMS-Auto dataset(See the above Setup section) to pretrain the SAFIRE image encoder.

Run the following command to start pretraining on multiple GPUs with DDP:

torchrun --nproc-per-node=6 pretrain.py --batch_size=2

🏗️ Train

We provide support for distributed data parallel (DDP) training using PyTorch. Below are the instructions to train the model using train.py:

Run the following command to start training on multiple GPUs with DDP:

torchrun --nproc-per-node=6 train.py --batch_size=6 --encresume="safire_encoder_pretrained.pth" --resume="" --num_epochs=150

Here are the explanations of the flags:

--nproc-per-node: Specifies the number of GPUs to use on a single node.
--batch_size: Sets the batch size per GPU. In this example, the total batch size is (6 * 6 = 36).
--encresume: Specifies the path to the pretrained encoder checkpoint file. It is uploaded to the Google Drive link provided in the Setup section.
--resume: Specifies the path to the model checkpoint file to resume training. Leave empty ("") to start training from scratch.
--num_epochs: Sets the total number of training epochs.

Make sure to adjust these parameters and paths in ForensicsEval/project_config.py.

©️ License

Update (2026-08-05): Working as a researcher in the industry made me realize that strict "research-only" licenses can actually make it quite difficult for corporate researchers to use open-source work, even for pure internal R&D. Because of this, I've decided to relax the licenses for the SAFIRE weights, as well as the SafireMS Dataset series, to CC-BY 4.0. And SAFIRE code to Apache 2.0.

This means you are now free to use them for commercial purposes, provided you give proper attribution.

Please Note: You are still responsible for checking and complying with the copyrights and licenses of any original data used to build the datasets, as well as any third-party code dependencies that SAFIRE relies on. Finally, I (Myung-Joon Kwon) assume no responsibility or liability for any issues arising from the use of the code or data in this repository.

📚 Citation

If you find this repository helpful, please cite our paper:

@article{kwon2024safire,
  title={SAFIRE: Segment Any Forged Image Region},
  author={Kwon, Myung-Joon and Lee, Wonjun and Nam, Seung-Hun and Son, Minji and Kim, Changick},
  journal={arXiv preprint arXiv:2412.08197},
  year={2024}
}

🔑 Keywords

SAFIRE, Segment Anything Model, SAM, Point Prompting, Promptable Segmentation, Image Forensics, Multimedia Forensics, Image Processing, Image Forgery Detection, Image Forgery Localization, Image Manipulation Detection, Image Manipulation Localization