Existing AIGI detectors learn generator-specific artifacts that become obsolete as generators evolve and break under real-world degradations. REM flips the paradigm: it models the real image distribution boundary, achieving robust detection that generalizes to unseen generators and survives chain degradations (social-media compression, filters, stickers, screenshots, ...).
We also release RealChain, a comprehensive benchmark with 7 state-of-the-art generators and 50 realistic degradation chains simulating real social media propagation.
- New Paradigm: Real-centric envelope modeling instead of artifact-driven detection — stable, transferable decision boundaries across quality domains
- State-of-the-Art: 94.9% average balanced accuracy across 12 benchmarks, +6.9% over previous SOTA
- Robustness: 84.2% on severely degraded RealChain (CD), outperforming existing methods by 18.4%
- Efficiency: Only requires real images + a VAE for boundary reconstruction — 10x lower GPU cost than diffusion-based data generation
REM consists of three key modules:
| Module | What it does |
|---|---|
| MBR (Manifold Boundary Reconstruction) | Injects controlled perturbations in VAE latent space to generate diverse near-real samples around the real manifold |
| EE (Envelope Estimator) | Learns a compact, smooth boundary via binary classification + tangency regularization |
| CDC (Cross-Domain Consistency) | Uses frozen DINO as anchor to keep the envelope stable across different quality domains |
RealChain is a comprehensive AIGI detection benchmark designed for realistic evaluation under real-world conditions.
| Category | Sources | Count |
|---|---|---|
| Real | MSCOCO, OpenImage-v7, Unsplash, ImageNet | 7,000 |
| Flux.1 (Open-source) | Text-to-Image | 1,000 |
| SDv3.5 (Open-source) | Text-to-Image | 1,000 |
| QwenImage (Open-source) | Text-to-Image | 1,000 |
| Hunyuan 3.0 (Commercial) | Text-to-Image | 1,000 |
| NanoBanana (Commercial) | Text-to-Image | 1,000 |
| Seedream 4.0 (Commercial) | Text-to-Image | 1,000 |
| Seedream 4.0 i2i (Commercial) | Image-to-Image | 1,000 |
Each image undergoes a randomly constructed degradation chain (length 2-5) simulating real social media propagation:
Propagation (cross-platform upload/download):
WeChat | TikTok | Baidu | Instagram | X (Twitter)
Post-processing (user editing):
Filter | Sticker | Crop/Resize | Screenshot
50 unique degradation chains are applied to both real and synthetic images, producing No Degradation (ND) and Chain Degradation (CD) versions.
# Via Hugging Face
git lfs install
git clone https://huggingface.co/datasets/handsomerich/RealChainRealChain/
├── Real/ # 7,000 real images
├── Flux1/ # 1,000 Flux.1 generated
├── SDv3.5/ # 1,000 SD v3.5 generated
├── QwenImage/ # 1,000 QwenImage generated
├── Hunyuan3/ # 1,000 Hunyuan 3.0 generated
├── NanoBanana/ # 1,000 NanoBanana generated
├── Seedream4/ # 1,000 Seedream 4.0 (t2i)
├── i2i/ # 1,000 Seedream 4.0 (i2i)
└── degradation_chains.json # 50 chain definitions
Each column shows original (ND) and degraded (CD) versions across all generators and real images. Chain degradations introduce JPEG artifacts, resolution loss, stickers, and color shifts — faithfully reproducing real social media environments.
- [2026/03] RealChain dataset is now open-sourced on HuggingFace!
- Training and inference code will be released soon. Stay tuned!
@article{liu2025beyond,
title={Beyond Artifacts: Real-Centric Envelope Modeling for Reliable AI-Generated Image Detection},
author={Liu, Ruiqi and Han, Yi and Zhang, Zhengbo and Yao, Liwei and Yan, Zhiyuan and Shen, Jialiang and Chen, ZhiJin and Sun, Boyi and Weng, Lubin and Dong, Jing and others},
journal={arXiv preprint arXiv:2512.20937},
year={2025}
}- Code: MIT License
- RealChain Dataset: CC BY-NC 4.0
The real images in RealChain are sourced from MSCOCO, OpenImage-v7, Unsplash, and ImageNet. Synthetic images are generated using open-source models (Flux.1, SDv3.5, QwenImage) and commercial APIs (Hunyuan 3.0, NanoBanana, Seedream 4.0).