Research map / Image generation / Editing and style transfer
Detective SAM: Adaptive AI-Image Forgery Localization
Gert Lek, Nicolas van Schaik, Chaoyi Zhu, Pin-Yu Chen, Robert Birke, Lydia Y. Chen
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 |
|---|---|---|---|
| add Gaussian blur and noise perturbation clues versus none (ablation) | Detective SAM · SIDA+MagicBrush validation | IoU: 36.22 -> 50.52 | Improved |
| add JPEG to Noise + Blur perturbations (ablation) | Detective SAM · SIDA+MagicBrush validation | IoU: 50.52 -> 52.58 | Improved |
| add Gaussian noise to Gaussian blur (ablation) | Detective SAM · SIDA+MagicBrush validation | IoU: 48.17 -> 50.52 | Improved |
| use transformer, uncertainty and spatial-gating mask adapter instead of simple convolution (ablation) | Detective SAM · SIDA+MagicBrush validation | IoU: 44.48 -> 50.52 | Improved |
| add feature adapters (ablation) | Detective SAM · SIDA+MagicBrush validation | IoU: 14.29 -> 50.52 | Improved |
| use SAM2 mask decoder instead of heatmap directly (ablation) | Detective SAM · SIDA+MagicBrush validation | IoU: 36.41 -> 50.52 | Improved |
| fine-tune on 500 FLUX-Bench and QWEN-Bench samples | Detective SAM · FLUX-Bench | IoU: 18.70 -> 43.08 | Improved |
| fine-tune on 500 FLUX-Bench and QWEN-Bench samples | Detective SAM · Avg OOD IoU | IoU: 34.68 -> 35.57 | Improved |
| fine-tune on 500 FLUX-Bench and QWEN-Bench samples | Detective SAM · NanoBanana | IoU: 19.34 -> 27.00 | Improved |
| add 20% replay during fine-tuning (ablation) | Detective SAM fine-tuning · ID and OOD IoU | IoU | Improved |
| fine-tune sequentially FLUX then QWEN versus FLUX only | Detective SAM · Avg OOD IoU | IoU: 35.90 -> 37.68 | Improved |
| overrepresent QWEN (1500) vs balanced fine-tuning | Detective SAM · QWEN-Bench / FLUX-Bench | IoU: 41.44 -> 44.94 (QWEN); 43.08 -> 42.16 (FLUX) | Mixed |
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.
- SIDA-7B — SIDA: Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model (CVPR 2025) · better on 3 of 3 shared results
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
- ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.86)
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.86)
- A Lottery Ticket Hypothesis Approach with Sparse Fine-tuning and MAE for Image Forgery Detection and Localization (AAAI 2025 · similarity 0.85)
- Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images (CVPR 2026 · similarity 0.85)
- Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis (NeurIPS 2025 · similarity 0.85)
- Less Supervision, Better Generalization: Weakly Supervised Fake Region Localization in Diffusion-Edited Images (NeurIPS 2026 · 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.
