Research map / Image generation / Autoregressive and latent generation
Few-Shot Learner Generalizes Across AI-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.
Reframes AI-generated image detection as few-shot classification, training a Prototypical Network with ResNet-50 to classify images by source generator, enabling detection of unseen generators with 10 samples.
Problem
Fake image detectors trained on synthetic datasets generalize poorly to unseen generative models, and collecting training data from closed-source generators is expensive or infeasible.
Approach
Treat each generator's outputs as a separate class and learn a metric space with a Prototypical Network so that unseen generators can be detected by nearest-prototype matching using only a few support samples (and zero-shot via metadata prototypes from training data).
Key results
- FSD with 10 shots reaches 84.1% average accuracy on GenImage unseen test subsets, versus 72.5% for LARE2 (+11.6%).
- FSD zero-shot achieves 77.1% average accuracy using metadata prototypes from training classes.
- Going from 1 to 10 shots raises ADM accuracy from 62.6% to 79.2%; 200 shots gives 81.7%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 10 support samples from the test class rather than no test-class samples | FSD · GenImage average | accuracy: 77.1 -> 84.1 | Helped |
| Use 10 support samples from the test class rather than no test-class samples | FSD · VQDM | accuracy: 69.1 -> 76.2 | Helped |
| Use 10 support samples from the test class rather than no test-class samples | FSD · BigGAN | accuracy: 62.1 -> 82.2 | Helped |
| Use 10 support samples rather than 1 support sample (ablation) | FSD · ADM | accuracy: 62.6% -> 79.2% | Helped |
| Increase support samples from 10 to 200 (ablation) | FSD · ADM | accuracy: 79.2% -> 81.7% | Helped |
| Use generator-specific classes rather than a single fake class (ablation) | FSD · ADM | accuracy | Helped |
Surprising findings
- FSD performs poorly on VQDM when it is excluded from training (due to image tokenization differences), indicating difficulty generalizing to vastly different generators.
- A binary real-vs-fake classifier fails to cluster unseen ADM samples, whereas per-generator multi-class training yields tighter clusters.
Limitations
- Performance declines when detecting images from generators vastly different from training generators (e.g., VQDM).
- Cannot obtain few-shot samples from all generative models, so training set covers only part of them.
- Distinguishing similar models such as SD v1.4 and v1.5 is difficult.
Experimental setup
Datasets: GenImage
Benchmarks: GenImage
Metrics: accuracy average precision
Baselines: Spec CNNSpot F3Net GramNet DIRE LARE2
Builds on: Prototypical Network CNNSpot DIRE LARE2
Closest papers by meaning
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.89)
- Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025 · similarity 0.87)
- WildFake: A Large-Scale and Hierarchical Dataset for AI-Generated Images Detection (AAAI 2025 · similarity 0.87)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.87)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 2025 · similarity 0.87)
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (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.