Research map / Label-efficient and robust learning / Domain adaptation and generalization
Smoothness Similarity Regularization for Few-Shot GAN Adaptation
Vadim Sushko, Ruyu Wang, Juergen Gall
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 smoothness similarity regularization L_SS to the generator (without L_all) (ablation) | StyleGANv2 few-shot adaptation · Face->Anime, Church->Shells | diversity (LPIPS) and quality | Mixed |
| add multi-layer discriminator loss L_all (without L_SS) (ablation) | StyleGANv2 few-shot adaptation · Face->Anime, Church->Shells | image quality and diversity | Mixed |
| combine L_SS and L_all (ablation) | StyleGANv2 few-shot adaptation · Face->Anime, Church->Shells | FID and LPIPS | Improved |
| replace PPL and MixDL regularizers with L_SS (ablation) | StyleGANv2 with L_all · Face->Anime, Church->Shells | FID and LPIPS | Improved |
| use L_all instead of PatchGAN discriminator loss (ablation) | StyleGANv2 discriminator · Church->Shells and Face->Anime | FID | Improved |
| increase lambda_SS to large values (25.0, 125.0) (ablation) | L_SS · Face->Anime, Church->Shells | FID vs LPIPS | Mixed |
| apply L_SS on very high-resolution generator features (e.g. 128x128) instead of 32x32 (ablation) | L_SS · Face->Anime, Church->Shells | image quality | Hurt |
| weight earlier D layers higher in L_all instead of uniform (ablation) | L_all · Face->Anime vs Church->Shells | FID | Mixed |
| weight later D layers higher in L_all instead of uniform (ablation) | L_all · Face->Anime, Church->Shells | FID vs LPIPS | Mixed |
| regularize smoothness w.r.t. class embeddings as well as noise in BigGAN (ablation) | BigGAN few-shot adaptation · ImageNet->Flowers, ImageNet->Pokemons | diversity (LPIPS) | Improved |
| apply method to structurally close source-target pairs (Face->Sketch, Face->Sunglasses) | Ours vs prior methods · Face->Sketch, Face->Sunglasses | FID and LPIPS | No clear effect |
Closest papers by meaning
- Generalized One-shot Domain Adaptation of Generative Adversarial Networks (NeurIPS 2022 · similarity 0.89)
- A Closer Look at Few-Shot Image Generation (CVPR 2022 · similarity 0.89)
- One-Shot Generative Domain Adaptation (ICCV 2023 · similarity 0.89)
- Few-shot Image Generation via Adaptation-Aware Kernel Modulation (NeurIPS 2022 · similarity 0.89)
- NICE: NoIse-modulated Consistency rEgularization for Data-Efficient GANs (NeurIPS 2023 · similarity 0.89)
- One-Shot Adaptation of GAN in Just One CLIP (TPAMI 2023 · similarity 0.88)
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.
