Research map / Image generation / Autoregressive and latent generation
WildFake: A Large-Scale and Hierarchical Dataset for AI-Generated Images 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.
Introduces WildFake, a large hierarchical dataset (about 3.57M images) of real and AI-generated images collected from open-source communities and generation pipelines, and benchmarks detectors' generalization and robustness on it.
Problem
Existing AI-generated image detection datasets cover only one or two generator types, fixed image categories, or low-quality user-generated images, limiting detector generalization and robustness evaluation.
Approach
A hierarchical dataset organized into cross-generator, cross-architecture, cross-weight, cross-time, and cross-version levels, with fake images sourced from open communities (Civitai, Midjourney) and personalized/fine-tuned models, enabling fine-grained generalization analysis.
Key results
- Detectors trained on WildFake reach the highest average accuracy/AP/AUC across test datasets, e.g., LASTED-WildFake averages 93.0/95.6/93.7 vs. 82.4/91.1/84.7 for LASTED-ArtiFact.
- Training ResNet50 on half of WildFake still outperforms GenImage, ArtiFact, and DiffusionDB trained on full data (average 83.4/91.5/87.5 at 1/2 vs. 81.9/90.0/80.7 for GenImage).
- ViT outperforms ResNet50 on degraded images, especially under geometric and color transformations; models trained on Midjourney and SD generalize best across DM architectures.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train detectors on WildFake rather than the strongest alternative training dataset for each detector | ResNet50, ViT, MultiLID, DIRE, IFDL, and LASTED · Average across testing datasets | AUC: 80.7 -> 89.42; 84.4 -> 91.1; 57.4 -> 61.8; 76.4 -> 86.6; 83.0 -> 99.0; 84.6 -> 93.7 | Helped |
| Increase WildFake training data from 1/16 to 1/8 (ablation) | ResNet50 · Average across testing datasets | AUC: 78.1 -> 81.3 | Helped |
| Increase WildFake training data from 1/8 to 1/4 (ablation) | ResNet50 · Average across testing datasets | AUC: 81.3 -> 84.6 | Helped |
| Increase WildFake training data from 1/4 to 1/2 (ablation) | ResNet50 · Average across testing datasets | AUC: 84.6 -> 87.5 | Helped |
| Increase WildFake training data from 1/2 to the full dataset (ablation) | ResNet50 · Average across testing datasets | AUC: 87.5 -> 89.4 | Helped |
Surprising findings
- Detectors trained on DMs generalize worse to GANs and Others than GANs/Others generalize to DMs, suggesting DMs' multi-step denoising creates a larger domain gap.
- Even at one-half of WildFake's volume, the dataset outperforms full-size baseline datasets, indicating data quality matters beyond volume.
Limitations
- Real-image and fake-image split is random 4:1 per generator, with no held-out generator-level split beyond the hierarchy experiments.
- Some fake images are sourced from community platforms of unverified generation settings, so generator provenance may be imperfect.
Experimental setup
Datasets: WildFake DiffusionDB ArtiFact GenImage DiffusionForensics MPBench COCO FFHQ ImageNet LSUN Church CelebA-HQ AFHQ
Benchmarks: WildFake DiffusionForensics GenImage DiffusionDB ArtiFact
Metrics: accuracy average precision AUC
Baselines: DiffusionDB ArtiFact GenImage DiffusionForensics MPBench DIRE IFDL multiLID LASTED ResNet50 ViT
Builds on: GenImage DIRE LASTED
Closest papers by meaning
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.89)
- Few-Shot Learner Generalizes Across AI-Generated Image Detection (ICML 2025 · similarity 0.87)
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.87)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.87)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.86)
- From Specificity to Generality: Revisiting Generalizable Artifacts in Detecting Face Deepfakes (NeurIPS 2025 · similarity 0.86)
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.