Research map / Image generation / GANs
Collapse by Conditioning: Training Class-conditional GANs 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 |
|---|---|---|---|
| Transition from unconditional to conditional training | StyleGAN2+ADA · ImageNet Carnivores, Food101, CUB-200-2011, AnimalFace | FID: 23 -> 14; 24 -> 20; 27 -> 22; 20 -> 16 | Helped |
| Transition from unconditional to conditional training | StyleGAN2+ADA · ImageNet Carnivores, Food101, CUB-200-2011, AnimalFace | KID: 0.0093 -> 0.0021; 0.0071 -> 0.0045; 0.0059 -> 0.0032; 0.0048 -> 0.0018 | Helped |
| Add generator and objective transition rather than using no transition (ablation) | StyleGAN2+ADA · Food101 and ImageNet Carnivores | FID: 79 -> 20; 110 -> 14 | Helped |
| Add transition in the generator (ablation) | Modified StyleGAN2+ADA · Food101 and ImageNet Carnivores | FID: 79 -> 25; 110 -> 17 | Helped |
| Add transition in the objective while retaining generator transition (ablation) | Transition-only-in-generator variant · Food101 and ImageNet Carnivores | FID: 25 -> 20; 17 -> 14 | Helped |
| Transition only the objective, keeping the generator conditional from the beginning (ablation) | Modified StyleGAN2+ADA · Food101 and ImageNet Carnivores | FID: 20 -> 80; 14 -> 107 | Hurt |
| Start the transition at training start (Ts=0) instead of Ts=2k (ablation) | Ours · Food101 | FID: 20 -> 24 | Hurt |
| Vary transition start time among Ts=1k, 2k, 4k, and 6k (ablation) | Ours · Food101 and ImageNet Carnivores | FID: Food101: 22, 20, 21, 23; ImageNet Carnivores: 15, 14, 14, 15 | No clear effect |
| Vary transition end time among Te=3k, 4k, and 5k (ablation) | Ours · Food101 and ImageNet Carnivores | FID: Food101: 21, 20, 23; ImageNet Carnivores: 16, 14, 15 | No clear effect |
| Use the transition method without ADA (ablation) | StyleGAN2 · ImageNet Carnivores, 50 classes and 500 images per class | FID: 13 -> 8 | Helped |
| Keep the unconditional loss throughout training instead of down-weighting it as conditioning increases (ablation) | Ours · Carnivores, Food101, CUB-200-2011, AnimalFace | FID: 18 -> 14; 25 -> 20; 25 -> 22; 20 -> 16 | 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)
- Rebooting ACGAN: Auxiliary Classifier GANs with Stable Training (NeurIPS 2021 · similarity 0.90)
- GAN Compression: Efficient Architectures for Interactive Conditional GANs (CVPR 2020 · similarity 0.89)
- A Unified View of cGANs with and without Classifiers (NeurIPS 2021 · similarity 0.89)
- Diversity-Sensitive Conditional Generative Adversarial Networks (ICLR 2019 · similarity 0.89)
- SAN: Inducing Metrizability of GAN with Discriminative Normalized Linear Layer (ICLR 2024 · 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.