Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Zero-shot Detection of AI-Generated Image via RAW-RGB Alignment
Haiwei Wu, Fengpeng Li, Zhilin Tu, Yuanman Li, Xiong Li, Jiantao Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| align RGB features with RAW-derived pipeline graphs and RAW latents, trained only on real RAW-RGB pairs | zero-shot AIGC detector · Zero-shot clustering avg over GANs and DMs | NMI: .742 -> .964 vs MIB | Improved |
| use RAW-RGB alignment trace instead of baseline features for pair similarity | zero-shot AIGC detector · Feature similarity avg | AUC: .734 -> .925 vs MIB | Improved |
| fine-tune the alignment feature space on ProGAN only | zero-shot AIGC detector · ForenSynths cross-domain | AP: .953 -> .987 vs Effort | Improved |
| freeze the pre-trained ViT instead of training it (ablation) | RGB-Vision branch · Ablation, detection | AP: .909 -> .963 | Improved |
| replace standard cross-attention with Stratified Entropy Attention (ablation) | RGB-Vision branch · Ablation, detection | AP: .969 -> .987 | Improved |
| use ViT backbone instead of ConvNeXt, EfficientNet or SegFormer (ablation) | RGB-Vision branch · Ablation, detection | AP: ConvNeXt .882, SegFormer .875 -> ViT (frozen) .963 | Improved |
| use all ISP plus post-processing pipelines instead of ISP-only or post-only (ablation) | RAW-Graph branch · Ablation, detection | AP: .843 / .858 -> .885 | Improved |
| encode pipelines as directed weighted graph instead of text or undirected graph (ablation) | RAW-Graph branch · Ablation, detection | AP: .885 text -> .917 undirected -> .958 directed -> .987 weighted | Improved |
| use InfoNCE instead of Triplet or Circle loss (ablation) | RAW-Graph branch · Ablation, detection | AP: .924 / .965 -> .987 | Improved |
| add RAW-Vision branch (ablation) | Alignment trace model · Ablation, detection | AP: .896 -> .987 | Improved |
| use VAE with KL constraint instead of UNet with L2 (ablation) | RAW-Vision branch · Ablation, detection | AP: .908 -> .975 | Improved |
| add cross-attention conditioning on trace feature in the RAW-Vision branch (ablation) | RAW-Vision branch · Ablation, detection | AP: .975 -> .987 | Improved |
Closest papers by meaning
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.88)
- Transfer Learning of Real Image Features with Soft Contrastive Loss for Fake Image Detection (AAAI 2025 · similarity 0.87)
- Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable (NeurIPS 2025 · similarity 0.86)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.86)
- Forensic Self-Descriptions Are All You Need for Zero-Shot Detection, Open-Set Source Attribution, and Clustering of AI-generated Images (CVPR 2025 · similarity 0.86)
- Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation (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.
