Research map / Label-efficient and robust learning / Domain adaptation and generalization
Few-shot Image Generation with Elastic Weight Consolidation
Yijun Li, Richard Zhang, Jingwan (Cynthia) Lu, Eli Shechtman
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 Fisher-weighted EWC while adapting the pretrained generator | StyleGAN few-shot image generation · 10-shot target domains | FID: 86.44~$\pm$~4.38 -> 74.87~$\pm$~3.75 | Improved |
| Use Fisher-weighted EWC while adapting the pretrained generator | StyleGAN few-shot image generation · 10-shot target domains | LPIPS diversity: 0.21~$\pm$~0.02 -> 0.40~$\pm$~0.02 | Improved |
| Use Fisher-weighted EWC while adapting the pretrained generator | StyleGAN few-shot image generation · 10-shot target domains | User fooling rate (%): 30.56~$\pm$~7.14 -> 47.92~$\pm$~7.38 | Improved |
| Use EWC regularization instead of setting λ=0 (ablation) | StyleGAN adapted to the target domain · 10-shot Fernand Léger faces | overfitting to training examples | Improved |
| Remove EWC regularization by setting λ=0 (ablation) | StyleGAN adapted to the target domain · 10-shot Fernand Léger faces | diversity of generated results | Hurt |
| Increase the number of target examples | Ours (EWC) · 10-shot to 10,000-shot target domains | FID: 74.87~$\pm$~3.75 -> 55.74~$\pm$~2.88 | Improved |
| Increase the number of target examples | BSA · 10-shot to 10,000-shot target domains | FID: 105.56~$\pm$~5.79 -> 131.20~$\pm$~6.04 | Hurt |
| Increase the number of target examples | NST · 1-shot to 10,000-shot target domains | FID: 212.23~$\pm$~9.77 -> 194.88~$\pm$~8.71 | No clear effect |
| Increase the number of target examples | MineGAN · 1-shot to 10,000-shot target domains | FID: 102.57~$\pm$~4.76 -> 58.69~$\pm$~3.14 | Improved |
| Choose a source domain more similar to the target | Pretrained generator adaptation · 10-shot cross-domain generation | FID | Improved |
| Use a larger regularization weight when source and target are more similar | EWC adaptation · Cross-domain generation | adaptation performance | Improved |
| Use a smaller regularization weight when more target data is available | EWC adaptation · Cross-domain generation | adaptation performance | Improved |
Closest papers by meaning
- One-Shot Domain Adaptation for Face Generation (CVPR 2020 · similarity 0.89)
- DomainGallery: Few-shot Domain-driven Image Generation by Attribute-centric Finetuning (NeurIPS 2024 · similarity 0.88)
- Few-shot Image Generation via Adaptation-Aware Kernel Modulation (NeurIPS 2022 · similarity 0.87)
- Few-shot Hybrid Domain Adaptation of Image Generator (ICLR 2024 · similarity 0.87)
- ImagineFSL: Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning (CVPR 2025 · similarity 0.86)
- MoLE: Enhancing Human-centric Text-to-image Diffusion via Mixture of Low-rank Experts (NeurIPS 2024 · 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.
