Research map / Detection and segmentation / Open-set and few-shot detection
Can We Build a Monolithic Model for Fake Image Detection? SICA: Semantic-Induced Constrained Adaptation for Unified-Yet-Discriminative Artifact Feature Space Reconstruction
Bo Du, Xiaochen Ma, Xuekang Zhu, Zhe Yang, Chaoqun Niu, Jian liu, Ji-Zhe 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 |
|---|---|---|---|
| replace full fine-tuning with SICA low-rank adaptation on CLIP | CLIP-FFT · OpenMMSec overall | accuracy: 80.2 -> 85.4 | Improved |
| replace Effort SVD-residual adaptation with SICA | Effort · OpenMMSec overall | accuracy: 80.1 -> 85.4 | Improved |
| use SICA instead of FFT under unified training | CLIP · Deepfake | AUC: 0.7900 -> 0.9234 | Improved |
| use SICA instead of FFT under unified training | CLIP · Doc | AUC: 0.8080 -> 0.8646 | Improved |
| unified training instead of single-domain training for FFT (ablation) | FFT · Deepfake | AUC: 0.8183 -> 0.7900 | Hurt |
| unified training instead of single-domain training for SICA (ablation) | SICA · Deepfake | AUC: 0.9235 -> 0.9234 | No clear effect |
| unified training instead of single-domain training for FFT (ablation) | FFT · AIGC | AUC: 0.9291 -> 0.8987 | Hurt |
| add leave-one-domain-out training data from other domains (ablation) | SICA · held-out Deepfake | AUC: 0.9235 single-domain vs 0.7660 LODO | Hurt |
| use CLIP-pretrained instead of ImageNet-1K ResNet-50 (ablation) | ResNet-50 · OpenMMSec overall | accuracy: 72.7 -> 78.3 | Improved |
| use CLIP ViT-L/14 instead of DINOv2 (ablation) | SICA · OpenMMSec overall | accuracy: 82.9 -> 85.4 | Improved |
| use CLIP ViT-L/14 instead of SigLIP (ablation) | SICA · OpenMMSec overall | accuracy: 82.2 -> 85.4 | Improved |
| increase LoRA rank beyond 8-32 (ablation) | SICA · OpenMMSec | accuracy: peak plateau at rank 8 to 32 | 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.
- SegFormer — SegFormer: A Topic Segmentation Model with Controllable Range of Attention (AAAI 2023) · 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
- ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.88)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.88)
- Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.87)
- BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors (NeurIPS 2026 · similarity 0.86)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (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.
