Research map / Image generation / Autoregressive and latent generation
HSIC Bottleneck for Cross-Generator and Domain-Incremental Synthetic Image Detection
Chin-Chia Yang, Yung-Yu Chuang, Hwann-Tzong Chen, Tyng-Luh Liu
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 intermediate ViT features (ablation) | HSIC bottleneck detector · Trained on SDV1.4 | mACC (%): 89.20 -> 92.22 | Improved |
| add intermediate ViT features (ablation) | HSIC bottleneck detector · Trained on ProGAN | mACC (%): 83.30 -> 89.22 | Improved |
| add HSIC(y,z) term (ablation) | HSIC bottleneck with intermediates · Trained on SDV1.4 | mACC (%): 92.22 -> 93.39 | Improved |
| add HSIC(x,z) term on top of HSIC(y,z) and intermediates (ablation) | HSIC bottleneck · Trained on SDV1.4 | mACC (%): 93.39 -> 93.86 | Improved |
| add HSIC(x,z) term on top of HSIC(y,z) and intermediates (ablation) | HSIC bottleneck · Trained on ProGAN | mACC (%): 90.67 -> 91.07 | Improved |
| add HSIC(x,z) alone (ablation) | CLIP detector · Trained on SDV1.4 | mACC (%): 89.20 -> 89.75 | No clear effect |
| use RBF median-bandwidth kernel instead of cosine (ablation) | HSIC bottleneck · Trained on SDV1.4 | mACC (%): 88.70 -> 93.86 | Improved |
| use RBF median-bandwidth kernel instead of RBF sigma=1 (ablation) | HSIC bottleneck · Trained on ProGAN | mACC (%): 88.82 -> 91.07 | Improved |
| add HSIC relevance to replay exemplar selection (ablation) | HGR · SDV1.4 start | mACC (%): 93.80 -> 94.38 | Improved |
| add HSIC relevance to replay exemplar selection (ablation) | HGR · ProGAN start | mACC (%): 86.23 -> 88.63 | Improved |
| use HGR replay instead of CBRS | domain-incremental detector · SDV1.4 start | mACC (%): 92.66 -> 94.38 | Improved |
| use CLIP ViT-L/14 instead of DINOv2 ViT-L/14 (ablation) | HSIC bottleneck · cross-generator detection | accuracy: 10 to 30 percentage points | Improved |
Closest papers by meaning
- SDDGR: Stable Diffusion-based Deep Generative Replay for Class Incremental Object Detection (CVPR 2024 · similarity 0.87)
- Revisiting Generative Replay for Class Incremental Object Detection (CVPR 2025 · similarity 0.87)
- FakeInversion: Learning to Detect Images from Unseen Text-to-Image Models by Inverting Stable Diffusion (CVPR 2024 · similarity 0.86)
- FerretNet: Efficient Synthetic Image Detection via Local Pixel Dependencies (NeurIPS 2025 · similarity 0.85)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.85)
- REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 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.
