Research map / Image generation / GANs
When Do GANs Replicate? On the Choice of Dataset Size
Qianli Feng, Chenqi Guo, Fabian Benitez-Quiroz, Aleix M. Martinez
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 one subset level (one-shot) instead of all levels to fit f1 | f1 exponential fit · BigGAN Flower | R2: 0.9739 -> 0.8228 | Hurt |
| use one subset level (one-shot) instead of all levels to fit f1 | f1 exponential fit · BigGAN CelebA | R2: 0.9388 -> 0.8301 | Hurt |
| use one subset level (one-shot) instead of all levels to fit f1 | f1 exponential fit · StyleGAN2 CelebA | R2: 0.9965 -> 0.9958 | No clear effect |
| use one subset level (one-shot) instead of all levels to fit f1 | f1 exponential fit · BigGAN LSUN-bedroom | R2: 0.8612 -> 0.8830 | Improved |
| add a second subset level (two-shot) to the one-shot fit | f1 exponential fit · BigGAN Flower | R2: 0.8228 -> 0.8230 | No clear effect |
| add a second subset level (two-shot) to the one-shot fit | f1 exponential fit · BigGAN CelebA | R2: 0.8301 -> 0.8174 | Hurt |
| use one-shot fit instead of full-subset fit | f1 exponential fit · BigGAN CelebA | MAE_f1 (%): 5.1180 -> 12.0495 | Hurt |
| use one-shot fit instead of full-subset fit (median over combinations) | f1 exponential fit · All combinations | MAE_f1 (%): 2.6840 -> 2.4215 | No clear effect |
| predict replication from dataset size via change of variable instead of from ID | f1/f2 fits · All combinations (one-shot) | MAE (%): 2.4215 -> 4.5558 | Hurt |
| use one-shot fit instead of full fit for predicting dataset size | f2 inverse · BigGAN LSUN-bedroom | MAE_f2^-1 (# samples): 2.4715e6 -> 1.4775e7 | Hurt |
| increase training dataset size | BigGAN and StyleGAN2 · CelebA, Flower, LSUN-bedroom | replication percentage: decay factor a in 0.96 to 0.98 | Improved |
Closest papers by meaning
- Are GANs Created Equal? A Large-Scale Study (NeurIPS 2018 · similarity 0.87)
- Consistency-GAN: Training GANs with Consistency Model (AAAI 2024 · similarity 0.87)
- RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 2024 · similarity 0.86)
- Instance Selection for GANs (NeurIPS 2020 · similarity 0.86)
- On the Statistical Properties of Generative Adversarial Models for Low Intrinsic Data Dimension (JMLR 2025 · similarity 0.86)
- CLR-GAN: Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction (ECCV 2024 · similarity 0.86)
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.
