Research map / Image generation / Autoregressive and latent generation
Attribution as Retrieval: Model-Agnostic AI-Generated Image Attribution
Hongsong Wang, Renxi Cheng, Chaolei Han, Jie Gui
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 |
|---|---|---|---|
| Use low-bit generative fingerprints instead of raw RGB input (ablation) | ResNet-50 · GenImage 1-shot attribution | mAP: +10.6% mAP | Improved |
| Add unsupervised pretext pre-training on real-image fingerprints (ablation) | LIDA · GenImage 1-shot attribution | mAP: +1.5% mAP | Improved |
| Add few-shot adaptation with center-based attribution loss (ablation) | LIDA · GenImage 1-shot attribution | mAP: +3.7% | Improved |
| Add real-prototype contrastive Deepfake detection loss (ablation) | LIDA · GenImage 1-shot attribution | mAP: +8.2% | Improved |
| Replace center loss LA with cross-entropy (ablation) | LIDA · GenImage 1-shot attribution | mAP: -1.8% | Hurt |
| Replace detection loss LD with cross-entropy (ablation) | LIDA · GenImage 1-shot attribution | mAP: -0.8% | Hurt |
| Replace both losses with cross-entropy (ablation) | LIDA · GenImage 1-shot attribution | mAP: -3.9% | Hurt |
| Increase detection loss weight lambda beyond 0.9 (ablation) | LIDA · GenImage attribution | Rank-1, mAP | Hurt |
| Increase registered shots from 1 to 10 per generator | LIDA · GenImage cross-architecture | Rank-1 vs mAP: Rank-1 +11.7% | Mixed |
| Use LIDA features instead of ResNet/DIRE/ESSP | LIDA · GenImage 10-shot attribution | Rank-1: +32.6%, +36.8%, +31.6% | Improved |
| Use LIDA instead of FSD for 10-shot detection | LIDA · GenImage detection | accuracy: 84.1 -> 88.3 | Improved |
| Merge diffusion generators into one paradigm class | LIDA · GenImage cross-generator | Rank-1, mAP: +30% Rank-1, +20% mAP | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AEROBLADE — AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error (CVPR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models (CVPR 2024 · 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)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.86)
- TraceEvader: Making DeepFakes More Untraceable via Evading the Forgery Model Attribution (AAAI 2024 · similarity 0.85)
- WildFake: A Large-Scale and Hierarchical Dataset for AI-Generated Images Detection (AAAI 2025 · similarity 0.85)
- Data Attribution for Text-to-Image Models by Unlearning Synthesized Images (NeurIPS 2024 · similarity 0.85)
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.
