Research map / Image generation / Autoregressive and latent generation
Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network
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.
Applies a variational information bottleneck on top of frozen CLIP ViT-L/14 image features (VIB-Net) to filter task-irrelevant information for universal AI-generated image detection, trained on ProGAN/ForenSynths or SD v1.4 and tested across 17-18 generators.
Problem
Detectors built on large pre-trained multimodal features (e.g., CLIP) carry many task-irrelevant features that induce superficial biases and hurt generalization to unseen generator families such as GANs versus diffusion models.
Approach
Insert a variational information bottleneck module after a frozen CLIP image encoder: a shallow MLP outputs a Gaussian latent z whose KL to a prior is penalized while a linear decoder predicts the real/fake label, so only task-relevant features reach the classifier.
Key results
- Trained on SD v1.4 (Genimage), VIB-Net improves average AP by 5.55% and ACC by 9.33% over SOTA methods across 15 test generators.
- On unseen GAN-family generators, VIB-Net improves AP by 12.48% and ACC by 23.59% over SOTA when trained on Genimage SD v1.4.
- Trained on ProGAN, VIB-Net improves AP by 2.26% and ACC by 6.19% over UnivFD on the ForenSynths test set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add VIB feature filtering to CLIP ViT-L/14 | CLIP-based generated-image detector · GenImage test (train on Stable Diffusion v1.4), Average | AP: 5.55% | Helped |
| Add VIB feature filtering to CLIP ViT-L/14 | CLIP-based generated-image detector · GenImage test (train on Stable Diffusion v1.4), Average | ACC: 9.33% | Helped |
| Add VIB feature filtering to CLIP ViT-L/14 | CLIP-based generated-image detector · Unseen generative models from different series | AP: 12.48% | Helped |
| Add VIB feature filtering to CLIP ViT-L/14 | CLIP-based generated-image detector · Unseen generative models from different series | ACC: 23.59% | Helped |
| Use CLIP ViT-L/14 rather than the other tested backbones (ablation) | VIB-Net · Stable Diffusion v1.4 training and generative-model tests | AP | Helped |
Surprising findings
- Diffusion-trained detectors (e.g., UnivFD, LGrad) drop about 40% in accuracy on ADM, Midjourney and VQDM, and models perform better on generators with distributions similar to the training data.
- CLIP features cluster by semantic category more strongly than by real/fake label, which the authors argue induces superficial biases.
Limitations
- Gains on the GAN-trained setting are modest relative to UnivFD, and performance on some generators (e.g., GauGAN, SAN) remains weaker.
- Classification accuracy depends on a fixed 0.5 threshold, which the authors note is sensitive to calibration.
- Evaluation limited to two datasets (ForenSynths and Genimage) and a restricted set of baselines.
Experimental setup
Datasets: ForenSynths Genimage LSUN ImageNet FaceForensics++ CelebA-HQ CelebA COCO
Benchmarks: ForenSynths Genimage
Metrics: average precision (AP) accuracy (ACC)
Baselines: CNNSpot FreDect Fusing LGrad DIRE UnivFD NPR CLIPping
Builds on: Information Bottleneck VIB UnivFD CLIP
Closest papers by meaning
- WildFake: A Large-Scale and Hierarchical Dataset for AI-Generated Images Detection (AAAI 2025 · similarity 0.85)
- Breaking Semantic Artifacts for Generalized AI-generated Image Detection (NeurIPS 2024 · similarity 0.83)
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.83)
- Beyond Generation: A Diffusion-based Low-level Feature Extractor for Detecting AI-generated Images (CVPR 2025 · similarity 0.83)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.83)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.83)
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.