ARForensics

# [ICCV 2025] D³QE: Learning Discrete Distribution Discrepancy-aware Quantization Error for Autoregressive-Generated Image Detection

<div align='center' style='margin-bottom:20px'>
<a href='http://arxiv.org/abs/2510.05891'><img src='https://img.shields.io/badge/ArXiv-red?logo=arxiv'></a> &nbsp;
<a href='https://ivg-yanranzhang.github.io/D3QE/'><img src='https://img.shields.io/badge/Visualization-green?logo=github'></a> &nbsp;
<a href="https://github.com/Zhangyr2022/D3QE"><img src="https://img.shields.io/badge/Code-9E95B7?logo=github"></a>
</div>

`ARForensics` dataset from [ICCV 2025] D³QE: Learning Discrete Distribution Discrepancy-aware Quantization Error for Autoregressive-Generated Image Detection

<div align=center>
<img src='dataset.png'>
</div>

## Introduction

The `ARForensics` dataset is the first benchmark for **visual autoregressive model detection**. It comprises 304,000 images (152,000 real from ImageNet, 152,000 synthetic). The dataset features 7 diverse autoregressive models: LlamaGen<small>[](https://github.com/FoundationVision/LlamaGen)</small>, VAR<small>[](https://github.com/FoundationVision/VAR)</small>, Infinity<small>[](https://github.com/FoundationVision/Infinity)</small>, Janus-Pro<small>[](https://github.com/deepseek-ai/Janus)</small>, RAR<small>[](https://github.com/bytedance/1d-tokenizer/tree/main)</small>, Switti<small>[](https://github.com/yandex-research/switti)</small>, and Open-MAGVIT2<small>[](https://github.com/TencentARC/SEED-Voken)</small>, which encompasses both token-based and scale-based architectures. It is structured into training (100k LlamaGen), validation (10k), and a comprehensive test set (6k samples from each of the 7 models), ensuring a balanced and technically varied evaluation benchmark for AI-generated image detection.

## Usage

After downloading all the files in this directory, you can run the following command to merge them into a single zip file:

```bash
cat ARForensics_part.* > ARForensics_part.zip
```

**Download:** The dataset `ARForensics` is uploaded and available at: [🤗 HuggingFace](https://huggingface.co/datasets/Yanran21/ARForensics) | [🤖 ModelScope](https://www.modelscope.cn/datasets/YanranZhang/ARForensics).

**Folder structure (expected):**

```
ARForensics/
├─ train/
│ ├─ 0_real/
│ └─ 1_fake/
├─ val/
│ ├─ 0_real/
│ └─ 1_fake/
└─ test/
├─ Infinity/
│ ├─ 0_real/
│ └─ 1_fake/
├─ Janus_Pro/
│ ├─ ..
├─ RAR/
├─ Switti/
├─ VAR/
├─ LlamaGen/
└─ Open_MAGVIT2/
```

## Cite

If you find this repository useful for your research, please consider citing this bibtex.

# [ICCV 2025] D³QE：面向自回归生成图像检测的离散分布差异感知量化误差学习

<div align='center' style='margin-bottom:20px'>
<a href='http://arxiv.org/abs/2510.05891'><img src='https://img.shields.io/badge/ArXiv-red?logo=arxiv'></a> &nbsp;
<a href='https://ivg-yanranzhang.github.io/D3QE/'><img src='https://img.shields.io/badge/Visualization-green?logo=github'></a> &nbsp;
<a href='https://github.com/Zhangyr2022/D3QE'><img src='https://img.shields.io/badge/Code-9E95B7?logo=github'></a>
</div>

`ARForensics` 数据集源自[ICCV 2025]论文《D³QE：面向自回归生成图像检测的离散分布差异感知量化误差学习》

<div align=center>
<img src='dataset.png'>
</div>

## 引言

`ARForensics` 数据集是首个面向**视觉自回归模型检测**的基准数据集。该数据集包含304,000张图像（其中152,000张来自ImageNet的真实图像，152,000张为合成图像）。数据集涵盖7种不同的自回归模型：LlamaGen<small>[](https://github.com/FoundationVision/LlamaGen)</small>, VAR<small>[](https://github.com/FoundationVision/VAR)</small>, Infinity<small>[](https://github.com/FoundationVision/Infinity)</small>, Janus-Pro<small>[](https://github.com/deepseek-ai/Janus)</small>, RAR<small>[](https://github.com/bytedance/1d-tokenizer/tree/main)</small>, Switti<small>[](https://github.com/yandex-research/switti)</small>以及Open-MAGVIT2<small>[](https://github.com/TencentARC/SEED-Voken)</small>，覆盖了基于Token（Token）和基于尺度的两类架构。数据集划分为训练集（10万张LlamaGen生成样本）、验证集（1万张样本）以及全面的测试集（7种模型各6000张样本），可为AI生成图像检测任务提供均衡且技术多样的评估基准。

## 使用方法

下载该目录下的所有文件后，可执行以下命令将其合并为单个压缩包：

bash
cat ARForensics_part.* > ARForensics_part.zip

**下载链接**：`ARForensics` 数据集已上传并可通过以下渠道获取：[🤗 HuggingFace](https://huggingface.co/datasets/Yanran21/ARForensics) | [🤖 ModelScope](https://www.modelscope.cn/datasets/YanranZhang/ARForensics)。

**预期文件夹结构**：

ARForensics/
├─ train/
│ ├─ 0_real/
│ └─ 1_fake/
├─ val/
│ ├─ 0_real/
│ └─ 1_fake/
└─ test/
├─ Infinity/
│ ├─ 0_real/
│ └─ 1_fake/
├─ Janus_Pro/
│ ├─ ..
├─ RAR/
├─ Switti/
├─ VAR/
├─ LlamaGen/
└─ Open_MAGVIT2/

## 引用

若本仓库对你的研究有所帮助，请考虑引用以下BibTeX条目。