README.md

---
language:
  - en
library_name: pytorch
tags:
  - image-classification
  - ai-generated-image-detection
  - deepfake-detection
  - universal-detector
  - robustness-asymmetry
  - dinov3
  - diffusion-generated-image-detection
license: other
pipeline_tag: image-classification
metrics:
  - accuracy
---

# RA-Det: Robustness Asymmetry Detection for AI-Generated Images

Official pretrained checkpoint for **RA-Det** _(ICML 2026)_ — a behavior-driven detector that decides whether an image is real or AI-generated by examining how it **behaves** under controlled perturbations, rather than how it **looks**.

> 📄 **Paper:** _RA-Det: Towards Universal Detection of AI-Generated Images via Robustness Asymmetry_ ([arXiv:2603.01544](https://arxiv.org/abs/2603.01544))
> 💻 **Code:** [github.com/dongdongunique/RA-Det](https://github.com/dongdongunique/RA-Det)
> 🏛️ **Venue:** ICML 2026

---

## TL;DR

Most detectors hunt for visual artifacts. Modern generators (Stable Diffusion, DALL·E, GANs) increasingly erase those artifacts, so appearance-based methods generalize poorly. RA-Det instead exploits a **robustness asymmetry**:

- **Natural images** keep a stable semantic representation under small perturbations.
- **AI-generated images** drift much more under the *same* perturbations.

RA-Det learns a perturbation operator that amplifies this gap, then classifies images from the resulting behavior — making it **data-agnostic, model-agnostic, and transferable to unseen generators**.

---

## How it works

RA-Det combines two components:

1. **Differential Robustness Probing (DRP)** — a learnable U-Net that applies a *bounded* perturbation (ε = 16/255) to the input, calibrated to maximally separate real vs. generated images in embedding space.
2. **Multi-Branch Detector** — aggregates three complementary cues:
   - **Semantic features** — from a frozen **DINOv3 ViT-L/16** foundation model (1024-d embeddings).
   - **Discrepancy features** — distance, similarity, and covariance statistics between the *clean* and *DRP-perturbed* embeddings (the core of the robustness-asymmetry signal).
   - **Low-level residual features** — capture high-frequency / forensic artifacts.

A **four-branch ensemble** variant adds an L²-based branch for additional robustness. This checkpoint is the recommended four-branch ensemble configuration.

> ⚠️ The statistics-based discrepancy branch needs enough samples per batch for stable covariance estimation. That is why **a batch size of 256 (32 per GPU × 8 GPUs) is strongly recommended** — smaller batches noticeably degrade accuracy.

---

## Model details

| | |
|---|---|
| **Configuration** | `ensemble_vitl16_raw_lpd_discrepancy` (four-branch ensemble) |
| **Backbone** | DINOv3 ViT-L/16, frozen (1024-d embeddings) |
| **Perturbation module** | Differential Robustness Probing (U-Net), ε = 16/255 |
| **Detector branches** | Semantic (DINOv3) + Discrepancy + Low-level residual (+ L²) |
| **Loss** | Discrepancy loss (margin = 1.0) + classification loss |
| **Training data** | ProGAN (Wang2020 format), `0_real/` & `1_fake/` per category |
| **Checkpoint file** | `checkpoint_best.pt` (PyTorch state_dict) |
| **Framework** | PyTorch (`torchrun` distributed training/eval) |
| **License** | See the [code repository](https://github.com/dongdongunique/RA-Det) (intended for research use) |

---

## Results

- Evaluated against **14+ diverse generative models** (ProGAN, Stable Diffusion v1.4, DALL·E 2, and more).
- Outperforms **10+ existing detection methods**.
- **+7.81% average performance improvement** over prior art, with strong transfer to **unseen generators**.

Per-generator accuracy tables and ablations are available in the [paper](https://arxiv.org/abs/2603.01544) and the [code repository](https://github.com/dongdongunique/RA-Det).

---

## How to use

This is **not** a single self-contained `.safetensors` model — RA-Det's forward pass requires the DINOv3 backbone plus the DRP U-Net and multi-branch detector defined in the codebase. The intended way to run it is via the official repository.

### 1. Get the code and weights

```bash
git clone https://github.com/dongdongunique/RA-Det.git
cd RA-Det
pip install -r requirements.txt

# DINOv3 backbone weights -> models/dino/dinov3_vitl16_pretrain.pth
# DINOv3 source -> models/dinov3_repo  (see repo README)
```

### 2. Download this checkpoint

```python
from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download(
    repo_id="dongdongunique/RA-Det-Checkpoints",
    filename="checkpoint_best.pt",
    repo_type="model",
)
print(ckpt_path)
```

### 3. Configure data paths

Edit `paths.py` (or set env vars) to point at your data in **Wang2020 format**:

```python
PROGAN_TRAIN_DATA_PATH = "/path/to/progan_train"   # .../airplane/{0_real,1_fake}/...
AIGCTEST_DATA_PATH     = "/path/to/AIGCTestset/test"  # progan/, stable_diffusion_v_1_4/, DALLE2/, ...
```

### 4. Run evaluation

The recommended entry point uses the same initialization as training:

```bash
bash scripts/validate.sh --gpus 8 --checkpoint /path/to/checkpoint_best.pt
```

…which expands to:

```bash
torchrun --nproc_per_node=8 train.py \
    --config ensemble_vitl16_raw_lpd_discrepancy \
    --eps 0.06274509803921569 \
    --margin 1.0 \
    --four-branch-ensemble \
    --normalize-loss \
    --validate-checkpoint /path/to/checkpoint_best.pt
```

### Reproduce training

```bash
# 8 GPUs, recommended batch size 256 (32/GPU), eps=16/255, margin=1.0
bash scripts/main.sh --gpus 8 --niter 1
```

| Parameter | Default | Note |
|---|---|---|
| `--eps` | 16/255 (0.0627) | Perturbation budget |
| `--margin` | 1.0 | Discrepancy-loss margin |
| `--niter` | 1 | Training epochs |
| `--gpus` | 8 | GPUs for `torchrun` |
| Batch size | **256** | Required for stable discrepancy statistics |

---

## Intended use & limitations

- **Intended for** research on AI-generated image detection, media forensics, and robustness analysis.
- **Out of scope:** real-world deployment as the sole decision in high-stakes moderation, surveillance, or identity-related judgments. Detector scores are probabilistic and can be wrong, especially on heavily post-processed, compressed, or adversarially crafted inputs.
- **Bias considerations:** detection performance varies across generators, image content, and demographics; treat outputs as one signal among several, never as ground truth.
- Generalization to **future, unseen generators** is a moving target — re-evaluate on your specific data before relying on results.

---

## Citation

If you use this checkpoint or the RA-Det method, please cite:

```bibtex
@misc{wang2026radetuniversaldetectionaigenerated,
      title={RA-Det: Towards Universal Detection of AI-Generated Images via Robustness Asymmetry},
      author={Xinchang Wang and Yunhao Chen and Yuechen Zhang and Congcong Bian and Zihao Guo and Xingjun Ma and Hui Li},
      year={2026},
      eprint={2603.01544},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.01544},
}
```