Research map / Image generation / GANs
On Leveraging Pretrained GANs for Generation with Limited Data
Read the paperCheck 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 |
|---|---|---|---|
| Transfer/freeze low-level generator and discriminator filters from pretrained GP-GAN (ablation) | GAN generation on CelebA · CelebA | FID: 19.48 -> 11.15 | Helped |
| Transfer/freeze low-level generator and discriminator filters from pretrained GP-GAN (ablation) | GAN generation on Flowers, Cars, and Cathedral · Flowers, Cars, Cathedral | training stability | Helped |
| Replace the original high-level part with the tailored compact SmallHead (ablation) | Transferred GP-GAN on Flowers, Cars, and Cathedral · Flowers, Cars, Cathedral | training stability | Helped |
| Add AdaFM modulation to the transferred filters (ablation) | SmallHead · CelebA | FID: 12.42 -> 9.90 | Helped |
| Add AdaFM modulation to the transferred filters (ablation) | SmallHead · Flowers | FID: 29.94 -> 16.76 | Helped |
| Add AdaFM modulation to the transferred filters (ablation) | SmallHead · Cars | FID: 20.64 -> 10.10 | Helped |
| Add AdaFM modulation to the transferred filters (ablation) | SmallHead · Cathedral | FID: 34.83 -> 15.78 | Helped |
| Transfer the G4D2 general part and use AdaFM rather than training the same architecture from scratch | GAN generation · CelebA | FID: 16.51 -> 9.90 | Helped |
| Transfer the G4D2 general part and use AdaFM rather than training the same architecture from scratch | GAN generation · Flowers | FID: 29.65 -> 16.76 | Helped |
| Transfer the G4D2 general part and use AdaFM rather than training the same architecture from scratch | GAN generation · Cars | FID: 11.77 -> 10.10 | Helped |
| Transfer the G4D2 general part and use AdaFM rather than training the same architecture from scratch | GAN generation · Cathedral | FID: 30.59 -> 15.78 | Helped |
| Transfer/freeze 3 lower discriminator groups instead of 2 (ablation) | Our-G4D2 · CelebA-1K | FID: 14.19 -> 13.99 | Helped |
Closest papers by meaning
- Training Generative Adversarial Networks with Limited Data (NeurIPS 2020 · similarity 0.92)
- On Stabilizing Generative Adversarial Training With Noise (CVPR 2019 · similarity 0.90)
- Constrained Generative Adversarial Networks for Interactive Image Generation (CVPR 2019 · similarity 0.90)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.89)
- RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 2024 · similarity 0.89)
- Diverse Rare Sample Generation with Pretrained GANs (AAAI 2025 · similarity 0.89)
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.