🪞 MIRROR

Manifold Ideal Reference ReconstructOR

A Reference-Comparison Framework for Generalizable AI-Generated Image Detection

🇨🇳 中文文档 · 📄 Paper · 📊 Results · 🚀 Quick Start · 📚 Citation

---

"Perception is a process of hypothesis testing."     Richard L. Gregory, 1980

MIRROR reframes AI-generated image (AIGI) detection: instead of binary artifact classification, it casts detection as a Reference-Comparison process. A learnable, orthogonal Memory Bank explicitly encodes the manifold of real images. Each input is projected onto this manifold via sparse top-$k$ attention to construct an Ideal Reference, and the comparison residual between the input and its reference becomes a generator-agnostic detection signal.

This shift unlocks two properties that prior detectors lack:

• 📈 Backbone scalability. Accuracy keeps climbing as DINOv3 scales from Base to Huge, while NPR / UnivFD / DDA saturate.
• 👁️ Superhuman robustness. On the human-imperceptible split of our Human-AIGI benchmark, MIRROR reaches 89.6% across 27 generators, surpassing both lay users and CV experts.

---

📰 News

• 2026.04  Inference code and DINOv3-H+ weights released.
• 2026.03  Paper released on arXiv.
• Coming soon  Training code, full checkpoint zoo, and the Human-AIGI Benchmark.

---

✨ Highlights

🔄 New paradigm	Reference-Comparison instead of artifact hunting
🏆 State of the art	+2.1% on 6 standard benchmarks, +8.1% on 7 in-the-wild benchmarks
👁️ Superhuman	89.6% on Human-AIGI hard subset, beats lay users and CV experts
📈 Scales with backbone	Sustained gains from DINOv3-Base to -Huge; competitors saturate
🧠 Reality memory bank	$K$ orthogonal prototypes encoding stable real-image regularities
🔬 First-of-its-kind benchmark	Psychophysically curated Human-AIGI set, 50 participants, 27 generators

---

🧭 How It Works

MIRROR is a two-phase framework (see fig/method.pdf).

Phase 1 · Encoding Reality Priors

A frozen DINOv3 encoder extracts patch-level features from real images only. A learnable memory bank $\mathbf{M} \in \mathbb{R}^{K \times D}$ of orthogonal prototypes is trained with sparse top-$k$ cross-attention reconstruction plus an orthogonality regularizer:

$$ \mathcal{L}_{\text{Phase1}} ;=; \lVert F - \hat{F} \rVert_2^2 ;+; \lambda ,\lVert \mathbf{M}\mathbf{M}^{\top} - \mathbf{I} \rVert_F $$

The first term forces $\hat{F}$ to faithfully reconstruct the real-image manifold; the second term keeps prototypes diverse and non-redundant.

Phase 2 · Reference-Comparison Detection

With $\mathbf{M}$ frozen, each input is projected onto the real-image manifold via the same sparse top-$k$ attention to produce its Ideal Reference $\hat{F}$. Real images align tightly with their reference; AI-generated images carry physical inconsistencies (illumination, geometry, texture statistics) that the reality memory cannot explain, producing a large comparison residual. The residual together with the reconstruction perplexity drive the final score.

---

📊 Results

All numbers are Balanced Accuracy (%) with format-aligned inputs (PNG to JPG).

13 standard + in-the-wild benchmarks

Category	Benchmark	Prior SOTA	DINOv2-L	DINOv3-L	DINOv3-H+	Δ vs SOTA
Standard	AIGCDetectBenchmark	84.7 (B-Free)	90.5	91.7	97.3	+12.6
	GenImage	89.6 (B-Free)	91.3	94.2	99.8	+10.2
	UnivFakeDetect	87.8 (B-Free)	84.6	88.2	92.4	+4.6
	Synthbuster	96.5 (DDA)	97.0	98.1	99.2	+2.7
	EvalGEN	96.6 (DDA)	98.1	99.0	99.8	+3.2
	DRCT-2M	99.2 (B-Free)	92.8	93.0	93.0	−6.2
In-the-wild	Chameleon	83.5 (DDA)	85.4	90.7	94.6	+11.1
	SynthWildx	94.6 (B-Free)	88.9	93.1	95.1	+0.5
	WildRF	92.6 (B-Free)	92.2	96.7	97.8	+5.2
	AIGIBench	84.4 (DDA)	85.6	90.5	94.9	+10.5
	CO-SPY	80.3 (DDA)	87.4	91.3	97.4	+17.1
	RR-Dataset	70.3 (DDA)	76.8	78.9	88.3	+18.0
	BFree-Online	87.1 (B-Free)	84.3	83.0	97.6	+10.5

Aggregate: +2.1% average on 6 standard benchmarks, +8.1% average on 7 in-the-wild benchmarks vs the previous SOTA.

Human-AIGI · the 14th benchmark

A psychophysically curated benchmark covering 27 generators, with a hard subset selected from a 50-participant study using accuracy, confidence, and response time. Designed to measure when detectors cross the Superhuman Crossover line.

Method	Hard subset Acc. (%)
Lay users (untrained)	~ 55
CV experts (no forensics training)	~ 73
MIRROR (DINOv3-H+)	89.6

See the paper for full psychophysics and the per-generator breakdown.

---

🛣️ Roadmap

• Inference code
• DINOv3-H+ inference weights
• Training code (Phase 1 + Phase 2)
• Full checkpoint zoo (DINOv2-L / DINOv3-L / DINOv3-H+)
• Human-AIGI Benchmark public release

---

🚀 Quick Start

1. Environment

Python 3.10+ is recommended.

git clone https://github.com/handsome-rich/MIRROR.git
cd MIRROR

# Install PyTorch first per your CUDA version: https://pytorch.org
pip install torch torchvision tqdm pillow numpy scikit-learn transformers peft

2. Download Weights

File	Purpose	Link
checkpoint-h-cur.pth	Phase 2 detector checkpoint	Google Drive
mirror_phase1.pth	Phase 1 memory-bank weights	Google Drive
dinov3-huge/	DINOv3-H+ backbone	official DINOv3 release

Place them under weight/:

weight/
├── checkpoint-h-cur.pth        # Phase 2 detector
├── mirror_phase1.pth           # Phase 1 memory bank
└── dinov3-huge/                # DINOv3-Huge backbone
    ├── config.json
    └── model.safetensors

3. Run Inference

python inference.py \
  --model_path     ./weight/checkpoint-h-cur.pth \
  --memory_path    ./weight/mirror_phase1.pth \
  --backbone_path  ./weight/dinov3-huge \
  --base_data_path /path/to/your/dataset \
  --benchmarks     Chameleon \
  --batch_size 128 \
  --device cuda \
  --use_amp

4. Dataset Layout

--base_data_path should point at a root that holds one folder per benchmark:

base_data_path/
├── AIGC_bm/                # AIGCDetectBenchmark
├── UniversalFakeDetect/    # UnivFD
├── synthbuster/            # Synthbuster
├── GenEval-JPEG/           # EvalGEN
├── Chameleon/test/
├── WildRF/test/
├── synthwildx/
├── AIGIBench/
├── CO-SPY-In-the-Wild/
├── drct/
├── RRDataset/
└── B-Free/

---

⚙️ Inference Arguments

Flag	Type	Description
--model_path	str	Phase 2 checkpoint (.pth)
--memory_path	str	Phase 1 memory-bank weights
--backbone_path	str	DINOv3 backbone directory
--base_data_path	str	Root directory containing benchmark sub-folders
--benchmarks	list	Benchmarks to evaluate, e.g. Chameleon GenImage
--batch_size	int	Per-device batch size
--device	str	cuda or cpu
--use_amp	flag	Enable mixed-precision inference
--output_dir	str	Where CSV reports go (default ./results)

CSV reports land at results/{benchmark}_{timestamp}.csv with Acc, Bal_Acc, Real_Acc, Fake_Acc.

---

📚 Citation

If MIRROR helps your research, please cite:

@article{liu2026mirror,
  title   = {MIRROR: Manifold Ideal Reference ReconstructOR for Generalizable AI-Generated Image Detection},
  author  = {Liu, Ruiqi and Cui, Manni and Qin, Ziheng and Yan, Zhiyuan and Chen, Ruoxin and Han, Yi and Li, Zhiheng and Chen, Junkai and Chen, ZhiJin and Lin, Kaiqing and others},
  journal = {arXiv preprint arXiv:2602.02222},
  year    = {2026}
}

---

📬 Contact

• Issues  github.com/handsome-rich/MIRROR/issues
• Email  ruiqi.liu24@nlpr.ia.ac.cn

---

_{🪞 Built on the conviction that understanding the real generalizes further than chasing the fake.}