Research map / Image generation / Autoregressive and latent generation
DGS-Net: Distillation-Guided Gradient Surgery for CLIP Fine-Tuning in AI-Generated Image Detection
Jiazhen Yan, Ziqiang Li, Fan Wang, Boyu Wang, Ziwen He, Zhangjie Fu
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 Orthogonal Suppression to LoRA-tuned CLIP (ablation) | CLIP-LoRA detector · AIGCDetectBench | mAcc: 79.9 -> 88.0 | Improved |
| add Prior Alignment to LoRA-tuned CLIP (ablation) | CLIP-LoRA detector · AIGCDetectBench | mAcc: 79.9 -> 83.3 | Improved |
| add Orthogonal Suppression to LoRA-tuned CLIP (ablation) | CLIP-LoRA detector · AIGIBench | average accuracy: 52.4 -> 61.7 | Improved |
| add Prior Alignment to LoRA-tuned CLIP (ablation) | CLIP-LoRA detector · AIGIBench | average accuracy: 52.4 -> 58.0 | Improved |
| combine Orthogonal Suppression and Prior Alignment (ablation) | CLIP-LoRA detector · AIGCDetectBench | mAcc: 79.9 -> 97.6 | Improved |
| combine Orthogonal Suppression and Prior Alignment (ablation) | CLIP-LoRA detector · AIGIBench | average accuracy: 52.4 -> 81.6 | Improved |
| naive LoRA fine-tuning of CLIP vs frozen CLIP, linear probing | CLIP image encoder · AIGIBench linear probing | accuracy: 71.5 -> 52.4 | Hurt |
| fine-tune with DGS-Net instead of naive LoRA, linear probing | CLIP image encoder · AIGIBench linear probing | accuracy: 52.4 -> 76.7 | Improved |
| fine-tune with DGS-Net instead of naive LoRA, linear probing | CLIP image encoder · AIGCDetectBench linear probing | accuracy: 82.4 -> 93.3 | Improved |
| DGS-Net vs retrained C2P-CLIP with same inference | CLIP-based detector · AIGCDetectBench | mAcc: +15.2% mAcc | Improved |
| DGS-Net under JPEG QF=75 | DGS-Net · AIGCDetectBench robustness | accuracy: 88.9 (QF=95) -> 82.4 (QF=75) | Mixed |
| Gaussian blur sigma 0.5 to 1.5 | DGS-Net · AIGCDetectBench robustness | accuracy: 92.7 -> 89.9 | No clear effect |
Closest papers by meaning
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.85)
- Improved Training Technique for Shortcut Models (NeurIPS 2025 · similarity 0.85)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.85)
- D^4: Dataset Distillation via Disentangled Diffusion Model (CVPR 2024 · similarity 0.85)
- Guided Score identity Distillation for Data-Free One-Step Text-to-Image Generation (ICLR 2025 · similarity 0.84)
- Breaking Latent Prior Bias in Detectors for Generalizable AIGC Image Detection (NeurIPS 2025 · similarity 0.84)
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.
