Research map / Image generation / GANs
Efficient Conditional GAN Transfer With Knowledge Propagation Across Classes
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 cross-class knowledge propagation and sharing instead of learning target BN parameters independently | BigGAN transfer · ImageNet to Places365 | FID: 85.9 -> 71.1 | Helped |
| use cross-class knowledge propagation and sharing instead of learning target BN parameters independently | BigGAN transfer · ImageNet to Places365 | KMMD: 0.18 -> 0.16 | Helped |
| use cross-class knowledge propagation | BigGAN transfer · CIFAR100 | FID: 28.03 -> 29.10; 30.89 -> 30.99; 40.73 -> 40.04; 30.73 -> 29.95; 35.62 -> 35.23; 56.48 -> 54.95 | Mixed |
| use cross-class knowledge propagation | BigGAN transfer · CIFAR100 | convergence speed: more than 4x speed-up on average | Helped |
| use cross-class knowledge propagation | BigGAN transfer · ImageNet to Places365 | convergence speed: more than 5x | Helped |
| add residual learning to a frozen prior (ablation) | BigGAN transfer · ImageNet to Places365 | FID: 96.3 -> 79.4 | Helped |
| make prior BN parameters tunable rather than frozen (ablation) | BigGAN transfer · ImageNet to Places365 | FID: 96.3 -> 81.2 | Helped |
| use layer-specific rather than shared similarity weights (ablation) | BigGAN transfer · ImageNet to Places365 | FID: 79.8 -> 81.2 | Helped |
| add both l1 similarity-score and l2 residual regularization (ablation) | BigGAN transfer · ImageNet to Places365 | FID: 82.4 -> 71.1 | Helped |
| add l1 regularization on similarity scores (ablation) | BigGAN transfer · ImageNet to Places365 | FID: 84.7 -> 71.1 | Helped |
| add l2 regularization on residuals (ablation) | BigGAN transfer · ImageNet to Places365 | FID: 81.1 -> 71.1 | Helped |
| fine-tune the rest of the model and residuals after learning target BN parameters | cGANTransfer · Places365 | FID: 71.1 -> 65.48 | Helped |
Closest papers by meaning
- Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions (CVPR 2024 · similarity 0.91)
- Diversity-Sensitive Conditional Generative Adversarial Networks (ICLR 2019 · similarity 0.89)
- MineGAN: Effective Knowledge Transfer From GANs to Target Domains With Few Images (CVPR 2020 · similarity 0.89)
- On Leveraging Pretrained GANs for Generation with Limited Data (ICML 2020 · similarity 0.89)
- Sequential Training of GANs Against GAN-Classifiers Reveals Correlated "Knowledge Gaps" Present Among Independently Trained GAN Instances (CVPR 2023 · similarity 0.89)
- SAN: Inducing Metrizability of GAN with Discriminative Normalized Linear Layer (ICLR 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.