Research map / Image generation / Image restoration
Deep Self-Dissimilarities as Powerful Visual Fingerprints
Idan Kligvasser, Tamar Rott Shaham, Yuval Bahat, Tomer Michaeli
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 |
|---|---|---|---|
| Use deeper VGG layers (ReLU51 vs ReLU21) for DSD2 full-reference | LDSD2 · PieAPP | Pearson correlation: 0.463 -> 0.761 | Improved |
| Use alpha=4 instead of alpha=2 in DSD full-reference measure | LDSD · PieAPP, ReLU41 | Pearson correlation: 0.769 -> 0.607 | Hurt |
| DSD instead of style loss | Full-reference measure · PieAPP, ReLU41 | Pearson correlation: 0.711 -> 0.769 | Improved |
| DSD instead of style loss at shallow layer | Full-reference measure · PieAPP, ReLU21 | Pearson correlation: 0.645 -> 0.463 | Hurt |
| Multi-scale style loss instead of single-scale style (ablation) | Style loss · PieAPP, ReLU31 | Pearson correlation: 0.714 -> 0.698 | Hurt |
| Use alpha=4 instead of alpha=2 in no-reference measure | QDSD · PieAPP no-reference | Pearson correlation: 0.330 -> 0.444 | Improved |
| DSD4 no-reference vs BRISQUE | QDSD4 · PieAPP no-reference | Pearson correlation: 0.363 -> 0.444 | Improved |
| Stop DSD minimization at 500 iterations instead of 5K | Image DSD optimization · ESRGAN output | DSD value / visual quality: 2.2 · 10−3 vs 1.2 · 10−3 | Mixed |
| Replace CX loss with DSD loss (ablation) | xSRResNet 5 blocks · DIV2K 4x SR, ReLU21 | NIQE: 7.87 -> 3.49 | Improved |
| Replace Style loss with DSD loss (ablation) | xSRResNet 5 blocks · DIV2K 4x SR, ReLU21 | NIQE: 3.64 -> 3.49 | Improved |
| Replace Style loss with DSD loss (ablation) | xSRResNet 5 blocks · DIV2K 4x SR, ReLU31 | NIQE: 4.20 -> 4.06 | Improved |
| Replace PDL with DSD loss (ablation) | xSRResNet 5 blocks · DIV2K 4x SR, ReLU22 | NIQE: 4.33 -> 3.78 | Improved |
Closest papers by meaning
- Attributing Fake Images to GANs: Learning and Analyzing GAN Fingerprints (ICCV 2019 · similarity 0.85)
- Minimal Images in Deep Neural Networks: Fragile Object Recognition in Natural Images (ICLR 2019 · similarity 0.84)
- Fingerprinting Deep Image Restoration Models (ICCV 2023 · similarity 0.84)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.84)
- Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment (CVPR 2025 · similarity 0.83)
- SeD: Semantic-Aware Discriminator for Image Super-Resolution (CVPR 2024 · similarity 0.83)
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.
