Ruoxin Chen1, Junwei Xi2, Zhiyuan Yan3, Keyue Zhang1, Shuang Wu1,
Jingyi Xie4, Xu Chen2, Lei Xu5, Isabel Guan6†, Taiping Yao1†, Shouhong Ding1
1Tencent YouTu Lab 2East China University of Science and Technology 3Peking University
4Renmin University of China 5Shenzhen University 6Hong Kong University of Science and Technology
2025/09: 🎉 Accepted by NeurIPS 2025 as Spotlight.2026/02: Official checkpoint released. We apologize for the previously uploaded checkpoint, whose performance slightly differed from the results reported in the paper.2026/03: Access information for 14 evaluation benchmarks has been released. We strongly recommend that future work in AIGI detection report results on all 14 benchmarks to better demonstrate overall performance and mitigate benchmark-specific bias.
In our DDA training pipeline, VAE-reconstructed fake images are compressed online using the same JPEG quality factor estimated from their paired real images. This step is critical for frequency-domain alignment between real and fake samples.
Specifically, our implementation applies online JPEG compression to VAE-reconstructed images with the same JPEG quality factor as their corresponding real images, with a probability of 0.5. We empirically find that applying this operation with a probability of 0.5 works better than always applying it with a probability of 1.0.
See Training/data/datasets.py, Lines 218–220:
if random.random() < 0.5:
jpeg_quality_factor = int(sample["jpeg_quality"])
fake_img = JPEG_Compression(fake_img, jpeg_quality_factor)JPEG compression with a quality factor of 96 is applied to the synthetic images in GenImage, ForenSynths, and AIGCDetectionBenchmark to mitigate format bias. The number of generators used in each dataset is reported: G refers to GAN, D to Diffusion, and AR to Auto-Regressive models. Among the 11 benchmarks, Chameleon, Synthwildx, WildRF, and Bfree-Online are the 4 in-the-wild datasets. Notably, DDA is the first detector to achieve over 80% cross-data accuracy on Chameleon.
| Benchmark | NPR (CVPR'24) | UnivFD (CVPR'23) | FatFormer (CVPR'24) | SAFE (KDD'25) | C2P-CLIP (AAAI'25) | AIDE (ICLR'25) | DRCT (ICML'24) | AlignedForensics (ICLR'25) | DDA (ours) |
|---|---|---|---|---|---|---|---|---|---|
| GenImage (1G + 7D) | 51.5 ± 6.3 | 64.1 ± 10.8 | 62.8 ± 10.4 | 50.3 ± 1.2 | 74.4 ± 8.4 | 61.2 ± 11.9 | 84.7 ± 2.7 | 79.0 ± 22.7 | 91.7 ± 7.8 |
| DRCT-2M (16D) | 37.3 ± 15.0 | 61.8 ± 8.9 | 52.2 ± 5.7 | 59.3 ± 19.2 | 59.2 ± 9.9 | 64.6 ± 11.8 | 90.5 ± 7.4 | 95.5 ± 6.1 | 98.1 ± 1.4 |
| DDA-COCO (5D) | 42.2 ± 5.4 | 52.4 ± 1.5 | 51.7 ± 1.5 | 49.9 ± 0.3 | 51.3 ± 0.6 | 50.0 ± 0.4 | 60.2 ± 4.3 | 86.5 ± 19.1 | 92.2 ± 10.6 |
| EvalGEN (3D + 2AR) | 2.9 ± 2.7 | 15.4 ± 14.2 | 45.6 ± 33.1 | 1.1 ± 0.6 | 38.9 ± 31.2 | 19.1 ± 11.1 | 77.8 ± 5.4 | 68.0 ± 20.7 | 97.2 ± 4.2 |
| Synthbuster (9D) | 50.0 ± 2.6 | 67.8 ± 14.4 | 56.1 ± 10.7 | 46.5 ± 20.8 | 68.5 ± 11.4 | 53.9 ± 18.6 | 84.8 ± 3.6 | 77.4 ± 25.0 | 90.1 ± 5.6 |
| ForenSynths (11G) | 47.9 ± 22.6 | 77.7 ± 16.1 | 90.0 ± 11.8 | 49.7 ± 2.7 | 92.0 ± 10.1 | 59.4 ± 24.6 | 73.9 ± 13.4 | 53.9 ± 7.1 | 81.4 ± 13.9 |
| AIGCDetectionBenchmark (7G + 10D) | 53.1 ± 12.2 | 72.5 ± 17.3 | 85.0 ± 14.9 | 50.3 ± 1.1 | 81.4 ± 15.6 | 63.6 ± 13.9 | 81.4 ± 12.2 | 66.6 ± 21.6 | 87.8 ± 12.6 |
| Chameleon (Unknown) | 59.9 | 50.7 | 51.2 | 59.2 | 51.1 | 63.1 | 56.6 | 71.0 | 82.4 |
| Synthwildx (3D) | 49.8 ± 10.0 | 52.3 ± 11.3 | 52.1 ± 8.2 | 49.1 ± 0.7 | 57.1 ± 4.2 | 48.8 ± 0.8 | 55.1 ± 1.8 | 78.8 ± 17.8 | 90.9 ± 3.1 |
| WildRF (Unknown) | 63.5 ± 13.6 | 55.3 ± 5.7 | 58.9 ± 8.0 | 57.2 ± 18.5 | 59.6 ± 7.7 | 58.4 ± 12.9 | 50.6 ± 3.5 | 80.1 ± 10.3 | 90.3 ± 3.5 |
| Bfree-Online (Unknown) | 49.5 | 49.0 | 50.0 | 50.5 | 50.0 | 53.1 | 55.7 | 68.5 | 95.1 |
| Avg ACC | 46.1 ± 16.1 | 56.3 ± 16.5 | 59.6 ± 14.6 | 47.6 ± 16.0 | 62.1 ± 15.6 | 54.1 ± 12.8 | 70.1 ± 14.6 | 75.0 ± 11.1 | 90.7 ± 5.3 |
| Min ACC | 2.9 | 15.4 | 45.6 | 1.1 | 38.9 | 19.1 | 50.6 | 53.9 | 81.4 |
All evaluation benchmarks used in our experiments are collected from publicly available sources.
We sincerely thank the original authors for making these valuable AIGI detection datasets publicly accessible.
Benchmarks marked with * are planned for future evaluation.
| Benchmark | Paper | Download |
|---|---|---|
| GenImage | GenImage: A Million-Scale Benchmark for Detecting AI-Generated Images | Google Drive |
| DRCT-2M | DRCT: Diffusion Reconstruction Contrastive Training towards Universe Detection of Diffusion Generated Images | ModelScope |
| Synthbuster | Synthbuster: Towards Detection of Diffusion Model Generated Images | Official Page |
| RAISE-1k (real subset used in Synthbuster) | Following the data source provided in ClipBased-SyntheticImageDetection | Official Download |
| ForenSynths | CNN-generated images are surprisingly easy to spot... for now | Google Drive · CMU Box |
| AIGCDetectionBenchmark | A Comprehensive Benchmark for AI-generated Image Detection | ModelScope |
| Chameleon | A Sanity Check for AI-generated Image Detection | Contact: tattoo.ysl@gmail.com |
| SynthwildX | Raising the Bar of AI-generated Image Detection with CLIP | GitHub |
| WildRF | Real-Time Deepfake Detection in the Real-World | Google Drive |
| Bfree-Online | A Bias-Free Training Paradigm for More General AI-generated Image Detection | Official Download |
| DDA-COCO | Dual Data Alignment Makes AI-Generated Image Detector Easier to Generalize | ModelScope |
| EvalGEN | Dual Data Alignment Makes AI-Generated Image Detector Easier to Generalize | Hugging Face |
| MNW* | Introducing the MNW Benchmark for AI Forensics | GitHub |
| AIGI-Now* | Task-Model Alignment: A Simple Path to Generalizable AI-Generated Image Detection | Hugging Face |
| T2I-CoReBench-Images* | Easier Painting Than Thinking: Can Text-to-Image Models Set the Stage, but Not Direct the Play? | Hugging Face |
The training dataset has been released on ModelScope and HuggingFace.
- New release: ModelScope
- Old ModelScope (legacy): ModelScope (deprecated / kept for reference)
DDA-COCO Benchmark has been released on ModelScope and HuggingFace.
EvalGEN Benchmark has been released on ModelScope and HuggingFace.
- Release arxiv paper with complete BibTeX citation
- Release checkpoint and inference code
- Release training set and training script
- Release code for DDA data construction
We sincerely thank @liyih for independently retraining the model using the official DDA training code and achieving improved performance compared to the results reported in the paper (see Issue #10). This validates the reproducibility of our implementation.
The reproduced checkpoint is available at: https://www.modelscope.cn/models/yihengli/DDA_reproduce
If you have any questions or suggestions, please feel free to contact us at cusmochen@tencent.com.
Feel free to reach out if you have any questions. This WeChat group brings together researchers working on AI-generated image detection, including authors of Effort (ICML 2025 Oral), DRCT (ICML 2024 Spotlight), FakeShield (ICLR 2025) and related work. Our goal is to build a focused community where researchers can exchange ideas and inspire new directions in AIGI detection.
Part of this codebase is adapted from UniversalFakeDetect. Huge thanks to the original authors for sharing their excellent work!
If you find this repository useful for your work, please consider citing it as follows:
@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}
}