Research map / Image generation / Autoregressive and latent generation
Detecting Generated Images by Fitting Natural Image Distributions
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 ConV, a training-free generated-image detector that checks whether a self-supervised model's (DINOv2) loss/representation is consistent under data-augmentation transforms along the natural manifold, plus F-ConV, which adds a RealNVP flow to push generated images off the natural manifold.
Problem
Generated image detectors trained as binary classifiers depend heavily on the quantity and diversity of generated training images and generalize poorly to images from unseen generative models.
Approach
Detect generated images by fitting only the natural image distribution: two functions (a self-supervised model's loss f1 and its composition with data augmentations f2 = f1∘h) are consistent on natural images but diverge on generated images because their gradients lie in orthogonal subspaces; F-ConV further uses a normalizing flow trained with a shaping and consistency loss to extrude generated im
Key results
- ConV (training-free, DINOv2 ViT-L/14) reaches 87.13 AUROC / 85.15 AP on ImageNet average, versus 93.68 / 93.11 for FatFormer (training-based).
- F-ConV reaches 93.77 AUROC / 93.38 AP on ImageNet average, the best among the compared methods, and 90.95 AUROC on the Sora/OpenSora average versus 89.36 for FatFormer.
- On GenImage (ACC), ConV averages 80.08 versus AEROBLADE's 73.83; F-ConV averages 90.1 ACC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the flow-based shaping and consistency losses (ablation) | F-ConV · ImageNet | AUROC: 89.17 -> 93.77 | Helped |
| Add the flow-based shaping and consistency losses (ablation) | F-ConV · ImageNet | AP: 87.89 -> 93.38 | Helped |
| Add the Shaping Loss (ablation) | F-ConV · ImageNet | AUROC: 89.17 -> 93.77 | Helped |
| Add the Consistency Loss (ablation) | F-ConV · ImageNet | AUROC: 92.54 -> 93.77 | Helped |
| Use HorizontalFlip in the transformation functions (ablation) | ConV · ImageNet | AUROC: 86.55 -> 87.13 | Helped |
| Use ColorJitter in the transformation functions (ablation) | ConV · ImageNet | AUROC: 85.97 -> 87.13 | Helped |
| Use GaussianBlur in the transformation functions (ablation) | ConV · ImageNet | AUROC: 85.49 -> 87.13 | Helped |
| Use random rotation (-45-45 degrees) instead of random rotation (-90-90 degrees) (ablation) | ConV · ImageNet average | AUROC: 71.34 -> 77.37 | Helped |
| Use random rotation at test time, which was not used during training (ablation) | ConV · ImageNet | AUROC: null | Hurt |
| Use CLIP instead of DINOv2 as the backbone (ablation) | ConV · ImageNet average | AUROC: 87.13 -> 72.99 | Hurt |
Surprising findings
- Using CLIP features with the same self-supervised consistency metric performs poorly; the authors needed a text-image similarity-based metric to recover performance.
- Random rotations not used in training augmentations do not yield good detection, showing dependence on transforms matching the self-supervised training augmentations.
Limitations
- No formal proof of generalization risk for the orthogonality-based approach.
- Performance drops on highly realistic generated images, where features remain nearly unchanged under transformation.
- Not evaluated against a generator adversarially trained to minimize the detector's inconsistency.
Experimental setup
Datasets: ImageNet LSUN-BEDROOM GenImage DRCT-2M ProGAN LAION
Benchmarks: ImageNet LSUN-BEDROOM GenImage DRCT-2M Sora/OpenSora frames
Metrics: AUROC AP ACC
Baselines: CNNspot Ojha DIRE NPR DRCT FatFormer AEROBLADE
Builds on: DINOv2 AEROBLADE RealNVP CNNspot DIRE
Closest papers by meaning
- Manifold Induced Biases for Zero-shot and Few-shot Detection of Generated Images (ICLR 2025 · similarity 0.91)
- Categorical Flow Matching on Statistical Manifolds (NeurIPS 2024 · similarity 0.86)
- What Secrets Do Your Manifolds Hold? Understanding the Local Geometry of Generative Models (ICLR 2025 · similarity 0.86)
- Anomaly Score: Evaluating Generative Models and Individual Generated Images based on Complexity and Vulnerability (CVPR 2024 · similarity 0.86)
- Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now (CVPR 2024 · similarity 0.86)
- Beyond Generation: A Diffusion-based Low-level Feature Extractor for Detecting AI-generated Images (CVPR 2025 · 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.