Research map / Image generation / Autoregressive and latent generation
Towards Universal Fake Image Detectors That Generalize Across Generative Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use nearest-neighbor classification on frozen CLIP ViT-L/14 features instead of the strongest trained-network baseline | Fake image detection trained with ProGAN data · Unseen diffusion and autoregressive models | mAP: +19.49 mAP | Helped |
| Use nearest-neighbor classification on frozen CLIP ViT-L/14 features instead of the trained-network baseline | Fake image detection trained with ProGAN data · Unseen diffusion and autoregressive models | accuracy: +25.90% acc | Helped |
| Train only a linear classifier on frozen CLIP ViT-L/14 features | Fake image detection trained with ProGAN data · Unseen diffusion and autoregressive models | mAP: +19.49 mAP | Helped |
| Replace ProGAN training data with LDM training data | CLIP feature linear classifier · Unseen GAN models | mAP: +37.16 mAP | Helped |
| Increase the nearest-neighbor pool from k=1 to k=9 (ablation) | CLIP ViT-L/14 nearest-neighbor detector · All generative models | average accuracy: 82.30 -> 84.25 | Helped |
| Use k=3 instead of k=1 neighbors (ablation) | CLIP ViT-L/14 nearest-neighbor detector · All generative models | mAP: 90.32 -> 90.22 | No clear effect |
| Use k=5 instead of k=1 neighbors (ablation) | CLIP ViT-L/14 nearest-neighbor detector · All generative models | mAP: 90.32 -> 90.22 | No clear effect |
| Use k=9 instead of k=1 neighbors (ablation) | CLIP ViT-L/14 nearest-neighbor detector · All generative models | mAP: 90.32 -> 90.14 | No clear effect |
| Use a CLIP-pretrained visual encoder rather than an ImageNet-pretrained encoder (ablation) | Linear classifier on frozen image features · Fake image detection across generative models | accuracy | Helped |
| Use CLIP ViT-L/14 rather than CLIP ResNet-50 (ablation) | Linear classifier on frozen image features · Fake image detection across generative models | accuracy | Helped |
| Use the first CLIP layer instead of a later layer for nearest-neighbor features (ablation) | CLIP nearest-neighbor detector · Fake image detection across generative models | accuracy | Hurt |
| Reduce training-set diversity from 20 classes to 2 classes (ablation) | CLIP nearest-neighbor detector · Fake image detection across generative models | accuracy | No clear effect |
Closest papers by meaning
- Few-Shot Learner Generalizes Across AI-Generated Image Detection (ICML 2025 · similarity 0.89)
- Transfer Learning of Real Image Features with Soft Contrastive Loss for Fake Image Detection (AAAI 2025 · similarity 0.88)
- FakeInversion: Learning to Detect Images from Unseen Text-to-Image Models by Inverting Stable Diffusion (CVPR 2024 · similarity 0.88)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.88)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.87)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.87)
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.