Research map / Image generation / Autoregressive and latent generation
ForensicConcept: Transferable Forensic Concepts for AIGI Detection
Menyanshu Zhou, Ziyin Zhou, Ke Sun, Yunpeng Luo, Jiayi Ji, Xiaoshuai Sun, Rongrong Ji
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 concept-aligned projection (concept branch) to DINOv3 detector (ablation) | DINOv3 detector · mean of GenImage, GAN-family, Chameleon | accuracy (%): 87.2 -> 88.8 | Improved |
| add concept branch to DINOv3 on GAN-family (ablation) | DINOv3 detector · GAN-family | accuracy (%): 87.3 -> 90.1 | Improved |
| add concept branch to DINOv3 on Chameleon (ablation) | DINOv3 detector · Chameleon | accuracy (%): 83.7 -> 84.4 | Improved |
| add concept branch to DINOv3 on GenImage (ablation) | DINOv3 detector · GenImage mean | accuracy (%): 90.7 -> 92.0 | Improved |
| inject CleanDIFT diffusion codebook into CLIP (ablation) | CLIP detector · GenImage mean | accuracy (%): 83.7 -> 88.2 | Improved |
| inject CleanDIFT codebook into CLIP on Midjourney (ablation) | CLIP detector · GenImage Midjourney | accuracy (%): 70.4 -> 85.9 | Improved |
| inject CleanDIFT codebook into CLIP on SDv1.4, SDv1.5, Wukong (ablation) | CLIP detector · GenImage SDv1.4 / SDv1.5 / Wukong | accuracy (%): -0.9, -1.1, -0.6 | Hurt |
| use diffusion-derived CleanDIFT codebook rather than other backbone codebooks for injection into CLIP (ablation) | CLIP detector · GenImage mean | accuracy (%): 88.2 vs 87.5 (EfficientNet), 85.4 (DINOv3) | Improved |
| use more CKNNA-aligned backbone as codebook source (ablation) | CLIP codebook injection · GenImage mean | accuracy (%) | Mixed |
| increase CKNNA neighborhood size kNN from 5 to 40 (ablation) | CKNNA alignment analysis · CleanDIFT us6 | CKNNA | Hurt |
| use ForensicConcept DINOv3 detector instead of CLIP or Effort under JPEG QF=75 | AIGI detector · GenImage with post-processing | accuracy (%): 63.37 -> 91.83 (vs Effort) | Improved |
Closest papers by meaning
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.88)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.86)
- Breaking Latent Prior Bias in Detectors for Generalizable AIGC Image Detection (NeurIPS 2025 · similarity 0.86)
- Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images (CVPR 2026 · similarity 0.85)
- BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors (NeurIPS 2026 · similarity 0.85)
- Forensic Self-Descriptions Are All You Need for Zero-Shot Detection, Open-Set Source Attribution, and Clustering of AI-generated Images (CVPR 2025 · similarity 0.85)
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.
