Fake2M: Benchmark for Fake-Image Detection
- Fake2M is a benchmark dataset for fake-image detection, comprising 3,174,000 images with 2,087,000 AI-generated fakes and 1,087,000 real photographs.
- It integrates diverse generative sources like Stable Diffusion, IF, and StyleGAN3 alongside real images from repositories such as CC3M, ensuring wide-ranging evaluation conditions.
- The dataset features rigorous quality controls, detailed metadata, and standardized tests (HPBench and MPBench) that reveal significant performance gaps between human (61.3%) and model (up to 87%) detection.
Fake2M is a large-scale visual benchmark dataset specifically constructed to support the training and evaluation of fake-image detectors in the presence of state-of-the-art AI-generated content. Created in the context of assessing both human and algorithmic performance at distinguishing authentic photographs from advanced synthetic imagery, Fake2M embodies extensive scale, diversity of generation sources, comprehensive labeling, and rigorous benchmarking protocols (Lu et al., 2023).
1. Dataset Construction and Composition
Fake2M comprises a total of 3,174,000 images, partitioned into 2,087,000 AI-generated (fake) images and 1,087,000 authentic (real) photographs. The dataset sources large and diverse generative models and real-world photography repositories.
Image Sources and Quantities
| Type | Source(s) | Quantity |
|---|---|---|
| Fake (diffusion) | Stable Diffusion v1.5 Realistic Vision V2.0 (SD-V1.5Real-dpms-25, 25-step DPM-Solver) | 1,000,000 |
| Fake (diffusion) | IF v1.0 with 25-step DPM-Solver++ (IF-V1.0-dpms++-25) | 1,000,000 |
| Fake (GAN) | StyleGAN3 variants (FFHQ, AFHQv2, MetFaces) | 87,000 |
| Real (captioned) | CC3M-Train (public Conceptual Captions repository, first 1M captions) | 1,000,000 |
| Real (faces) | StyleGAN3-Train (FFHQ, AFHQv2, MetFaces) | 87,000 |
The dataset maintains a fake-image fraction .
Each image is accompanied by detailed metadata, including generator type, textual prompt or caption (for text-to-image sets), sampling method and step count (for diffusion models), random seed, CFG-scale (for diffusion), resolution, and, for StyleGAN3 faces, matched real dataset labels.
2. Data Collection and Curation Protocols
Real photographs were obtained from the Conceptual Captions repository (CC3M) and, for the HPBench subset, augmented with images from 500px and Google Images using identical category prompts. Generative model outputs were systematically produced:
- Stable Diffusion and IF models utilized fixed random seeds and solver settings (25 inference steps) to maximize photorealism and reproducibility.
- StyleGAN3 images were generated from pre-trained checkpoints over controlled random seeds, targeting the FFHQ, AFHQv2, and MetFaces domains.
Manual quality control was implemented at multiple stages. Annotators filtered generated images to remove samples with egregious artifacts (e.g., extreme blurring, gross anatomical distortions), ensuring the dataset is populated with high-fidelity fakes. For HPBench, a further round excluded trivially-detectable fakes, focusing human evaluation on realistic, challenging examples.
3. Dataset Splits, Training, and Validation
Training and validation splits reflect diverse source combinations for robust benchmarking. Four “dataset settings” (A, B, C, D) are defined, varying the aggregations of fake and real subsets to probe model sensitivity to source diversity.
Validation is conducted using MPBench, a protocol built around 14 held-out sets covering both generator and real-image families:
- Diffusion models: SD-V2.1, SD-V1.5, SD-V1.5Real, and multiple IF sampling settings (10/25/50 steps, DPM-Solver++ / DDIM / DDPM; 15,000 images per set)
- Autoregressive model: CogView2 (22,000)
- Unknown commercial crawl: Midjourney (5,500)
- GAN: StyleGAN3 (60,000)
- Reals: ImageNet-Test (100,000), CelebA-HQ-Train (24,000), CC3M-Val (15,000)
Training is recommended on the full set across settings (A–D), with validation and reporting structured per-set and as an aggregate.
4. Benchmark Frameworks: HPBench and MPBench
HPBench (Human Perception Benchmark) and MPBench (Model Perception Benchmark) establish standardized evaluation protocols.
- HPBench: Fifty volunteers in a controlled laboratory environment each evaluated 100 randomly sampled images (50 real, 50 fake) with unlimited time, providing binary real/fake labels and, for fakes, categorizing the defect or selecting “intuition.” Human accuracy was 61.3% (misclassification rate, MR_human = 38.7%). Humans demonstrated greater proficiency in recalling real photos (67%) than in detecting fakes (56%).
- MPBench: Classification models (e.g., ConvNext, ResNet50, CLIP-ViT) were trained to distinguish real and fake, then evaluated on the 14 held-out sets, with key metrics including per-set accuracy, average accuracy by set type, and aggregate scores. The best model (ConvNext-S with Blur+JPEG 0.5 augmentation, Dataset Setting D) achieved 87% accuracy (MR_model = 13%) on HPBench and approximately 83% mean accuracy on MPBench. Protocols for HPBench comparability demand that models be evaluated under identical test splits as human trials.
5. Metadata, Labeling Schema, and Annotations
Every image in Fake2M is comprehensively annotated. Annotations encompass:
- Generator/model identifier (e.g., “SD-V1.5Real-dpms-25”)
- Associated prompt or caption from CC3M
- Denoising and inference settings (for diffusion models)
- Random seeds and configuration guidance scale (CFG)
- Output resolution (512×512, 256×256, etc.)
- For StyleGAN3, linkage to the originating real dataset subset (FFHQ, AFHQv2, MetFaces)
Such metadata is intended to facilitate fine-grained analysis of detector generalization across model families, domains, and inference hyperparameters.
6. Applications, Benchmarking, and Access
Fake2M is designed to underpin several lines of inquiry:
- Training and benchmarking of robust fake-image detectors with broad source coverage
- Comparative analysis of human and machine vulnerabilities to photorealistic fakes
- Generalization studies across diverse generator families (diffusion, GANs, autoregressive)
- Applied research in audit and forensics for journalism, law enforcement, and social media moderation
The dataset and associated HPBench/MPBench benchmarking code are publicly available under an open-source research license. Downloads and further documentation are hosted at https://github.com/Inf-imagine/Sentry. Proper attribution to “Seeing is not always believing: Benchmarking Human and Model Perception of AI-Generated Images” (Lu et al., 2023) is required in publications and derivative works.
7. Significance and Context
Fake2M establishes a new standard for rigor and scale in fake-image detection research. Its sampling, curation, and evaluation protocols enable a nuanced understanding of evolving threats posed by AI-generated images to the trustworthiness of photographic content. By combining a large, richly annotated corpus with standardized human and model benchmarks, it furnishes both a realistic challenge and a detailed measurement apparatus for progress and shortcomings in the field. The observed gap between human (61.3% accuracy) and leading model performance (87% on HPBench, ~83% across MPBench validation) demonstrates current limitations and the need for continued advances in reliable fake-image detection (Lu et al., 2023).
