Research map / Image generation / Autoregressive and latent generation
Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Dual Data Alignment (DDA), which aligns real and VAE-reconstructed synthetic images in both frequency (JPEG quality matching) and pixel (mixup) domains to train a DINOv2-LoRA AI-generated image detector on MSCOCO, and introduces DDA-COCO and EvalGEN test sets.
Problem
AI-generated image detectors overfit to spurious non-causal cues (format, size, semantics, frequency richness) in biased training data, so they fail to generalize to unbiased or unseen-generator test sets; pixel-level alignment via reconstruction is shown to remain misaligned in frequency.
Approach
Extends VAE-reconstruction data alignment with a frequency-level step (apply the real image's estimated JPEG quality factor to reconstructions) followed by pixel-level mixup between real and frequency-aligned synthetic images, so synthetic training samples are close to real ones in both domains.
Key results
- DDA reaches 90.7% average balanced accuracy across 11 benchmarks, about 15.7 points above the second-best method, with minimum accuracy 81.4%.
- DDA achieves 82.4% on Chameleon and 97.2% on EvalGEN, versus 71.0% and 77.8% for the strongest baselines.
- DDA trained with 118K/118K data needs about 5.9 h to build the training set versus 8.73 h for AlignedForensics and 64.6 h for DRCT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on dual pixel- and frequency-aligned MSCOCO pairs with DDA | DINOv2 detector · GenImage | balanced accuracy: 84.7 -> 91.7 | Helped |
| train on dual pixel- and frequency-aligned MSCOCO pairs with DDA | DINOv2 detector · EvalGEN | balanced accuracy: 77.8 -> 97.2 | Helped |
| train on dual pixel- and frequency-aligned MSCOCO pairs with DDA | DINOv2 detector · ForenSynths | balanced accuracy: 92.0 -> 81.4 | Hurt |
| train on dual pixel- and frequency-aligned MSCOCO pairs with DDA | DINOv2 detector · 11 benchmarks (average) | balanced accuracy: 75.0 -> 90.7 | Helped |
| use input sizes between 224 and 504 instead of 336 (ablation) | DDA detector · five-benchmark average | balanced accuracy: 88.2 ± 11.5; 88.4 ± 10.5; 88.4 ± 11.3; 89.3 ± 10.2; 87.1 ± 12.4; 88.2 ± 9.1; 91.9 ± 6.1 | No clear effect |
| use DINOv2 instead of CLIP as the LoRA-finetuned backbone (ablation) | DDA detector · five-benchmark average | balanced accuracy: 83.2 ± 14.4 -> 91.9 ± 6.1 | Helped |
| use LoRA fine-tuning instead of linear probing (ablation) | DDA detector with CLIP ViT-B/16 · five-benchmark average | balanced accuracy: 74.3 ± 18.5 -> 76.1 ± 21.5 | Helped |
| use linear probing on DINOv2 ViT-S/14 instead of DINOv2 LoRA fine-tuning (ablation) | DDA detector · five-benchmark average | balanced accuracy: 91.9 ± 6.1 -> 65.3 ± 6.2 | Hurt |
| use linear probing on DINOv2 ViT-B/14 instead of DINOv2 LoRA fine-tuning (ablation) | DDA detector · five-benchmark average | balanced accuracy: 91.9 ± 6.1 -> 66.8 ± 4.9 | Hurt |
| use linear probing on DINOv2 ViT-L/14 instead of DINOv2 LoRA fine-tuning (ablation) | DDA detector · five-benchmark average | balanced accuracy: 91.9 ± 6.1 -> 65.0 ± 7.8 | Hurt |
| use SD21 rather than other VAEs for training-data reconstruction (ablation) | DDA detector · DDA benchmarks | accuracy | Helped |
| set pixel-alignment probability and mixup upper bound to values between 0.2 and 0.8 rather than 0.0 or 1.0 (ablation) | DDA detector · DDA benchmarks | accuracy | Mixed |
Surprising findings
- Reconstructed images restore high-frequency detail that real JPEG images lack, so a frequency-based detector (SAFE) labels visually identical VAE reconstructions as synthetic about 93% of the time, indicating a format-driven shortcut.
- Several prior detectors (NPR, SAFE, AIDE, SAFE on DDA-COCO) score near chance, with large real-vs-fake accuracy gaps on DDA-COCO, suggesting learned dataset bias.
Limitations
- Performance on real-world images is limited by heavy post-processing and smartphone AI enhancements that can look synthetic-like.
- Underperforms on ForenSynths, where FatFormer and C2P-CLIP trained on ProGAN have an advantage and older small GAN generators deviate from the training distribution.
- Relies on a single VAE (SD 2.1) for training data reconstruction; the paper notes SD21 is best among the VAEs tested.
Experimental setup
Datasets: MSCOCO GenImage DRCT-2M Synthbuster AIGCDetectionBenchmark ForenSynths Chameleon WildRF SynthWildx BFree-Online GenEval prompts
Benchmarks: GenImage DRCT-2M DDA-COCO EvalGEN Synthbuster ForenSynths AIGCDetectionBenchmark Chameleon SynthWildx WildRF BFree-Online
Metrics: balanced accuracy average accuracy minimum accuracy standard deviation across subsets
Baselines: NPR UnivFD FatFormer SAFE C2P-CLIP AIDE DRCT AlignedForensics
Builds on: AlignedForensics DRCT LoRA DINOv2
Closest papers by meaning
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.88)
- Aligned Datasets Improve Detection of Latent Diffusion-Generated Images (ICLR 2025 · similarity 0.87)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.86)
- BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors (NeurIPS 2026 · similarity 0.85)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.85)
- FakeInversion: Learning to Detect Images from Unseen Text-to-Image Models by Inverting Stable Diffusion (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.