Research map / Image generation / GANs
Adaptive Feature Interpolation for Low-Shot Image Generation
Mengyu Dai, Haibin Hang, Xiaoyang Guo
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 adaptive feature interpolation (AFI) to FastGAN | FastGAN · Shells (1024x1024) | FID: 152.53 -> 124.80 | Improved |
| add AFI to StyleGAN2 hinge loss with n=20 output neurons | StyleGAN2 · Shells (1024x1024) | FID: 101.72 -> 62.99 | Improved |
| add AFI to StyleGAN2 hinge loss with n=20 output neurons | StyleGAN2 · Art (1024x1024) | FID: 58.42 -> 43.94 | Improved |
| use multidimensional discriminator output (n=20) in StyleGAN2 without feature interpolation (ablation) | StyleGAN2 · Shells (1024x1024) | FID: 123.66 -> 101.72 | Improved |
| add AFI to CIFAR-10 baseline with 100 training images | CIFAR-10 baseline · CIFAR-10 (100 images) | FID: 42.80 -> 27.62 | Improved |
| add AFI to CIFAR-10 baseline with 1% training data | CIFAR-10 baseline · CIFAR-10 (1% data) | FID: 19.69 -> 13.50 | Improved |
| mixup-style feature interpolation instead of AFI (ablation) | StyleGAN2 (no transformation augmentation) · Shells | FID: 229.83 -> 319.80 | Hurt |
| input-space mixup instead of AFI (ablation) | StyleGAN2 (no transformation augmentation) · Shells | FID: 229.83 -> 135.66 | Improved |
| use skewed Dirichlet weights (t=1) instead of uniform (t=0) for neighbor interpolation (ablation) | FastGAN + feature interpolation · Shells (k=2) | FID: 140.89 -> 130.98 | Improved |
| interpolate features with k=2 neighbors instead of no interpolation (k=1), uniform weights (ablation) | FastGAN + feature interpolation · Shells | FID: 165.72 -> 140.89 | Improved |
| use feature interpolation instead of image interpolation (mixup) in StyleGAN2 training (ablation) | StyleGAN2 · Shells (FID during training) | training stability | Improved |
| add AFI to StyleGAN2 hinge n=20 on Dog 256x256 vs Grumpy cat | StyleGAN2 · Dog (256x256) | FID: 56.43 -> 49.14 | Improved |
Closest papers by meaning
- Low-shot Learning via Covariance-Preserving Adversarial Augmentation Networks (NeurIPS 2018 · similarity 0.88)
- On Leveraging Pretrained GANs for Generation with Limited Data (ICML 2020 · similarity 0.88)
- F-VAEGAN-D2: A Feature Generating Framework for Any-Shot Learning (CVPR 2019 · similarity 0.87)
- Training Generative Adversarial Networks with Limited Data (NeurIPS 2020 · similarity 0.87)
- A Characteristic Function Approach to Deep Implicit Generative Modeling (CVPR 2020 · similarity 0.87)
- Improving Generative Adversarial Networks with Denoising Feature Matching (ICLR 2017 · 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.
