Research map / Image generation / Autoregressive and latent generation
Epistemic Uncertainty for Generated Image Detection
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 WePe, a training-free AI-generated image detector that flags images whose DINOv2 features change strongly under random weight perturbation (epistemic uncertainty proxy), plus a fine-tuned variant WePe* that amplifies the uncertainty gap.
Problem
Detecting AI-generated images that generalize to unseen generators, without training on generated images, given distribution shift between natural and generated images.
Approach
Use epistemic uncertainty of a large vision model pretrained only on natural images as a proxy for generated-image status, estimating it cheaply by weight perturbation (feature similarity between clean and perturbed DINOv2 is high for natural images, low for generated ones), with an optional fine-tuning step (WePe*) that widens the gap.
Key results
- WePe (training-free) reaches 85.04 average AP on ImageNet-based benchmark versus 81.58 for RIGID, per the ImageNet table (average AUROC 83.58 for RIGID vs 87.99 for WePe with Gaussian noise).
- WePe* achieves average AUROC 95.57 and AP 94.33 on ImageNet, the best among training and training-free methods.
- On GenImage, WePe (training-free) averages 80.84 ACC versus 78.19 for RIGID and 73.83 for AEROBLADE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune WePe to amplify the uncertainty gap (ablation) | WePe · ImageNet, average across generators | AUROC: 87.99 -> 95.57 | Helped |
| fine-tune WePe to amplify the uncertainty gap (ablation) | WePe · GenImage, average | ACC: 80.84 -> 93.2 | Helped |
| use weight perturbation instead of MC Dropout (ablation) | WePe · ImageNet | AUROC: 81.63 -> 87.99 | Helped |
| use uniform weight noise instead of Gaussian weight noise (ablation) | WePe · ImageNet | AUROC: 87.99 -> 89.06 | Helped |
| use Laplace weight noise instead of Gaussian weight noise (ablation) | WePe · ImageNet | AUROC: 87.99 -> 87.13 | Hurt |
| use DINOv2 ViT-S/14 instead of ViT-L/14 (ablation) | WePe · ImageNet | AUROC: 87.99 -> 72.83 | Hurt |
| use DINOv2 ViT-B/14 instead of ViT-L/14 (ablation) | WePe · ImageNet | AUROC: 87.99 -> 81.82 | Hurt |
| use DINOv2 ViT-g/14 instead of ViT-L/14 (ablation) | WePe · ImageNet | AUROC: 87.99 -> 84.92 | Hurt |
| use CLIP ViT-L/14 instead of DINOv2 ViT-L/14 (ablation) | WePe · ImageNet | AUROC: 87.99 -> 84.82 | Hurt |
| use BLIP instead of DINOv2 (ablation) | WePe · ImageNet | AUROC: 87.99 -> 68.25 | Hurt |
| use ImageNet classifier MSP or entropy scores instead of weight-perturbation uncertainty (ablation) | WePe · ImageNet | AUROC: 87.99 vs 48.85, 51.76, 63.23, 65.79 | Hurt |
| use multiple weight perturbations instead of a single perturbation (ablation) | WePe · ImageNet | AUROC | Helped |
Surprising findings
- MC Dropout performs worse than weight perturbation (AUROC 81.63 vs 87.99), since DINOv2 is not trained with dropout.
- Image classification models (ResNet18, ViT-L/16) using MSP or entropy scores perform near chance for generated-image detection.
Limitations
- Performance depends on a gap in uncertainty between natural and generated images that may narrow as generators improve.
- Out-of-distribution natural images such as cartoons are misclassified as generated.
- Weaker on CLIP and multimodal models (BLIP AUROC 68.25) than on DINOv2.
Experimental setup
Datasets: ImageNet LSUN-BEDROOM GenImage DRCT-2M
Benchmarks: ImageNet LSUN-BEDROOM GenImage DRCT-2M
Metrics: AUROC AP ACC
Baselines: RIGID AEROBLADE CNNspot Ojha DIRE PatchCraft FatFormer DRCT NPR
Builds on: MC Dropout Deep Ensembles RIGID UnivFD DINOv2
Closest papers by meaning
- Epistemic Uncertainty Quantification For Pre-Trained Neural Networks (CVPR 2024 · similarity 0.86)
- Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods? (ICML 2024 · similarity 0.85)
- From Risk to Uncertainty: Generating Predictive Uncertainty Measures via Bayesian Estimation (ICLR 2025 · similarity 0.85)
- Experts Don't Cheat: Learning What You Don't Know By Predicting Pairs (ICML 2024 · similarity 0.85)
- Evaluating Epistemic Uncertainty: Beyond OOD Detection and Active Learning (NeurIPS 2026 · similarity 0.85)
- Are Uncertainty Quantification Capabilities of Evidential Deep Learning a Mirage? (NeurIPS 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.