README.md

---
license: apache-2.0
library_name: pytorch
pipeline_tag: image-classification
tags:
  - aigc-detection
  - ai-generated-image-detection
  - image-forensics
  - pytorch
  - dinov3
  - clip
  - icml-2026
---

<div align="center">

# 🔍 ForensicConcept

### Transferable Forensic Concepts for AIGI Detection

Official detector checkpoints for the ICML 2026 paper
**ForensicConcept: Transferable Forensic Concepts for AIGI Detection**.

[[`Code`](https://github.com/EthanAdamm/FORENSICCONCEPT)]
[[`Paper`](https://arxiv.org/abs/2606.07034)]
[[`Models`](https://huggingface.co/ethan225/ForensicConcept/tree/main/weights)]
[[`BibTeX`](#citation)]

[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://www.apache.org/licenses/LICENSE-2.0)
[![PyTorch](https://img.shields.io/badge/PyTorch-2.7%2B-ee4c2c.svg)](https://pytorch.org/)
[![ICML 2026](https://img.shields.io/badge/ICML-2026-8A2BE2.svg)](https://icml.cc/)

</div>

---

## Model Overview

ForensicConcept converts diffuse detector evidence into explicit, auditable
forensic concepts and transfers these concepts across detector backbones. The
framework combines:

- **Adapter-guided discriminative tuning (ADT)** for efficient detector
  adaptation with LoRA;
- **Unsupervised concept induction (UCI)** for discovering a compact forensic
  concept codebook from decision-critical image patches;
- **Concept-aligned projection (CAP)** for concept-space prediction;
- **Concept-guided codebook injection (CGCI)** for transferring
  diffusion-derived generation traces to target backbones.

<div align="center">
  <img src="https://raw.githubusercontent.com/EthanAdamm/FORENSICCONCEPT/main/assets/readme/method_overview.png" alt="ForensicConcept overview" width="95%" />
</div>

## Performance

Image-level accuracy (%) reported in the paper. Models are trained on Stable
Diffusion 1.4 images.

| Evaluation benchmark | DINOv3 without concepts | ForensicConcept |
|---|---:|---:|
| GenImage, mean over 8 generators | 90.7 | **92.0** |
| GAN-family, mean over 7 generators | 87.3 | **90.1** |
| Chameleon | 83.7 | **84.4** |
| Mean over the three benchmarks | 87.2 | **88.8** |

The transferred codebook also raises CLIP ViT-L/14 mean accuracy on GenImage
from **83.7** to **88.2** (+4.5 points).

## Available Files

| File | Purpose | Size |
|---|---|---:|
| `weights/detectors/dinov3_vitl16_lora_stage1.pth` | DINOv3 ADT / LoRA initialization | 24.3 MiB |
| `weights/detectors/dinov3_vitl16_concept_stage2.pth` | DINOv3 ForensicConcept inference | 26.6 MiB |
| `weights/detectors/clip_vitl14_lora_stage1.pth` | CLIP LoRA initialization | 4.6 MiB |
| `weights/detectors/clip_vitl14_codebook_stage2.pth` | CLIP CGCI inference | 20.5 MiB |
| `weights/concepts/dinov3_concept_matrix.npy` | DINOv3 concepts (`200 x 1024`, float32) | 0.8 MiB |

The four detector checkpoints total **76.0 MiB**. Stage-1 checkpoints are
provided for reproducing stage-2 training. For inference, use the
corresponding stage-2 checkpoint. DINOv3 inference also requires the released
concept matrix.

## Quick Start

### 1. Clone the code

```bash
git clone https://github.com/EthanAdamm/FORENSICCONCEPT.git
cd FORENSICCONCEPT
python -m pip install -r requirements.txt
```

### 2. Download the released weights

Run this command from the root of the cloned code repository:

```bash
hf download ethan225/ForensicConcept \
  --include "weights/**" \
  --local-dir .
```

### 3. Add the external backbones

The pretrained DINOv3 ViT-L/16 and CLIP ViT-L/14 backbones are not included.
Download them from their official upstream projects and use this layout:

```text
weights/
|-- backbones/
|   |-- ViT-L-14.pt
|   `-- dinov3-vitl16/
|       `-- model.safetensors
|-- concepts/
|   `-- dinov3_concept_matrix.npy
`-- detectors/
    |-- dinov3_vitl16_lora_stage1.pth
    |-- dinov3_vitl16_concept_stage2.pth
    |-- clip_vitl14_lora_stage1.pth
    `-- clip_vitl14_codebook_stage2.pth
```

### 4. Evaluate

```bash
# DINOv3 ForensicConcept
python test_with_config.py \
  --config configs/dinov3_concept.yaml \
  --checkpoint_path weights/detectors/dinov3_vitl16_concept_stage2.pth

# CLIP with concept-guided codebook injection
python test_with_config.py \
  --config configs/clip_codebook.yaml \
  --checkpoint_path weights/detectors/clip_vitl14_codebook_stage2.pth
```

Dataset paths are placeholders in the public configurations. Update
`data.*` and `testing.groups` before evaluation. See the
[`GitHub README`](https://github.com/EthanAdamm/FORENSICCONCEPT#dataset-preparation)
for complete installation, dataset, training, and evaluation instructions.

## Checkpoint Notes

- DINOv3 stage 2 contains the classifier, concept mapping/head, and all LoRA
  tensors. The external concept matrix initializes the model structure before
  loading the checkpoint.
- CLIP stage 2 contains visual LoRA, the main classifier, and the complete
  codebook head.
- External backbone weights remain subject to their respective upstream
  licenses and terms.

<details>
<summary><strong>SHA-256 checksums</strong></summary>

```text
3bc8e833d75c9ccbb214c28238c791681294ae8b0c50e48cc3e64b9e5ac5ca1f  weights/detectors/dinov3_vitl16_lora_stage1.pth
99162dbf6a56610be5a8bb9fa27e62311f723d0a4f34e384b084e76085491aad  weights/detectors/dinov3_vitl16_concept_stage2.pth
b0c0217b547391a70eed680ef0d8f92555d172241c46bea767346c8d65115184  weights/detectors/clip_vitl14_lora_stage1.pth
a81ebec9281aca3db4b0d88254d25424f37d5d498ae2c405eb6a5be4748dc498  weights/detectors/clip_vitl14_codebook_stage2.pth
9e6ce224d223dab804648bb46a9cc405bfef4cfcb13dc9cc21ad7fd94935959c  weights/concepts/dinov3_concept_matrix.npy
```

</details>

## Citation

```bibtex
@inproceedings{zhou2026forensicconcept,
  title     = {{ForensicConcept}: Transferable Forensic Concepts for {AIGI} Detection},
  author    = {Zhou, Menyanshu and Zhou, Ziyin and Sun, Ke and Luo, Yunpeng and Ji, Jiayi and Sun, Xiaoshuai and Ji, Rongrong},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026}
}
```

## License

The ForensicConcept detector checkpoints are released under the
[Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0). External
backbone weights are not included and remain subject to their respective
upstream licenses and terms.