---
title: EvalGEN
canonical_url: "https://www.modelscope.cn/datasets/JunweiXi/EvalGEN"
md_url: "https://www.modelscope.cn/datasets/JunweiXi/EvalGEN.md"
repository: JunweiXi/EvalGEN
last_updated: 2026-10-19
license: "Apache License 2.0"
storage_size: "14 GB"
downloads: 5727
stars: 0
---

# EvalGEN

> EvalGEN - JunweiXi 在 ModelScope 开源的数据集。Official Dataset for Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable Conference: 39th Conference on Neural Information Processing Systems (NeurIPS 2025) https://arxiv.org/abs/2505.14359

JunweiXi/EvalGEN 是 ModelScope 魔搭社区上的数据集，存储大小 14 GB，采用 Apache License 2.0 许可。

- **Repository**: JunweiXi/EvalGEN
- **License**: Apache License 2.0
- **Storage size**: 14 GB
- **Downloads**: 5727
- **Stars**: 0
- **Last updated**: 2026-10-19

Source: https://www.modelscope.cn/datasets/JunweiXi/EvalGEN

---

# EvalGEN Benchmark

### Official Dataset for **Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable**
**Conference:** 39th Conference on Neural Information Processing Systems (NeurIPS 2025) https://arxiv.org/abs/2505.14359

---

#### Dataset Description
**EvalGEN** is a challenging, future-oriented AIGI detection benchmark designed to test detector generalization against the **latest generative models from 2024-2025** and **heterogeneous architectures** (e.g., auto-regressive models).

Unlike traditional benchmarks based solely on Stable Diffusion or GANs, EvalGEN introduces the latest **Auto-Regressive** generators and multimodal large models, covering high resolution, strong semantic reasoning, and unified generation frameworks.

#### Details
* **Prompt Source:** 553 distinct aligned prompts from the GenEval Benchmark.
* **Total Size:** Approximately 55,300 images (11,060 images per generator).
* **Included Generators (5 Latest Models):**
    1.  **Flux:** SOTA-level diffusion model with superior image quality.
    2.  **GoT (Generative of Things):** Multimodal model combining LLM reasoning with diffusion processes.
    3.  **Infinity:** High-resolution generative model based on Bitwise Auto-regressive modeling.
    4.  **OmniGen:** Unified multimodal generation framework.
    5.  **NOVA:** Non-quantized auto-regressive model balancing high fidelity and efficiency.
* **Format:** Unified as JPEG (Quality Factor 96) to simulate real-world web environments.

#### Dataset Note

During a recent review of EvalGEN, we identified a small number of missing and duplicate images. The missing images are unavailable in the original data and cannot be recovered. The duplicate images have been removed from the dataset.

#### Acknowledgments

We sincerely thank [@wj-sim](https://github.com/wj-sim) for identifying these issues and helping us improve the quality of EvalGEN. We appreciate their careful examination and valuable feedback. Please refer to [GitHub Issue #15](https://github.com/roy-ch/Dual-Data-Alignment/issues/15) for further details.

#### Citation
```code
@inproceedings{chen2025dual,
  title={Dual Data Alignment Makes {AI}-Generated Image Detector Easier Generalizable},
  author={Ruoxin Chen and Junwei Xi and Zhiyuan Yan and Ke-Yue Zhang and Shuang Wu and Jingyi Xie and Xu Chen and Lei Xu and Isabel Guan and Taiping Yao and Shouhong Ding},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025},
  url={https://openreview.net/forum?id=C39ShJwtD5}
}
```