Research map / Label-efficient and robust learning / Knowledge distillation
Semantic Relation Preserving Knowledge Distillation for Image-to-Image Translation
Zeqi Li, Ruowei Jiang, Parham Aarabi
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 2 direction semantic relation preserving loss to vanilla KD (ablation) | CycleGAN student · horse->zebra | FID: 106.10 -> 86.31 | Improved |
| add 2 direction SP loss to vanilla KD (ablation) | CycleGAN student · apple->orange | FID: 127.21 -> 121.17 | Improved |
| add 2 direction SP loss to vanilla KD (ablation) | CycleGAN student · tiger->leopard | FID: 82.04 -> 81.17 | Improved |
| add 2 direction SP loss to vanilla KD (ablation) | CycleGAN student · leopard->tiger | FID: 87.29 -> 80.75 | Improved |
| add 2 direction SP loss to vanilla KD (ablation) | CycleGAN student · summer->winter | FID: 80.10 -> 76.59 | No clear effect |
| apply SP loss in both generators rather than only the first (ablation) | CycleGAN student · horse->zebra | FID: 90.65 -> 86.31 | Improved |
| add intermediate KD loss on first fake image (ablation) | CycleGAN student · horse->zebra | FID: 106.10 -> 97.20 (vs +SP 90.65) | Hurt |
| add 2 direction SP loss to vanilla KD | CycleGAN student · Cityscapes label->photo | Mean pixel accuracy: 0.584 -> 0.704 | Improved |
| add SP loss at layer 3 of UNet to vanilla KD (ablation) | Pix2Pix UNet256 student · Cityscapes label->photo | Mean pixel accuracy: 0.742 -> 0.770 | Improved |
| distill SP at layer 2 instead of layer 3 (ablation) | Pix2Pix UNet256 student · Cityscapes label->photo | Mean pixel accuracy: 0.770 -> 0.743 | Hurt |
| add 2 direction SP loss to vanilla KD (ablation) | ResNet6 CycleGAN student · horse->zebra | FID: 105.49 -> 105.51 | No clear effect |
| add 2 direction SP loss to vanilla KD (ablation) | ResNet6 CycleGAN student · zebra->horse | FID: 146.26 -> 141.90 | Improved |
Closest papers by meaning
- Distilling Portable Generative Adversarial Networks for Image Translation (AAAI 2020 · similarity 0.91)
- Relation-Guided Adversarial Learning for Data-Free Knowledge Transfer (IJCV 2024 · similarity 0.90)
- Wavelet Knowledge Distillation: Towards Efficient Image-to-Image Translation (CVPR 2022 · similarity 0.90)
- E$^2$GAN: Efficient Training of Efficient GANs for Image-to-Image Translation (ICML 2024 · similarity 0.87)
- Unlocking Pre-trained Image Backbones for Semantic Image Synthesis (CVPR 2024 · similarity 0.87)
- Compressing Image-to-Image Translation GANs Using Local Density Structures on Their Learned Manifold (AAAI 2024 · similarity 0.87)
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.
