Research map / Image generation / GANs
Ensembling Off-the-Shelf Models for GAN Training
Check 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 |
|---|---|---|---|
| Add progressively selected frozen pretrained feature discriminators alongside the original discriminator | StyleGAN2 · Full-data LSUN Cat, LSUN Church, and LSUN Horse | FID: 6.86 -> 3.98; 4.28 -> 1.72; 4.09 -> 2.11 | Helped |
| Add vision-aided discriminators | StyleGAN2 on FFHQ · FFHQ-1024 full dataset | FID: 2.98 -> 3.01 | No clear effect |
| Add vision-aided discriminators | StyleGAN2 on FFHQ · FFHQ-1024 full dataset | PPL: 144.62 -> 127.58 | Helped |
| Use best linear-probe model selection instead of random selection (ablation) | Progressive vision-aided training · FFHQ 1k | FID: 15.48 -> 11.43 | Helped |
| Use best linear-probe model selection instead of least-accurate model selection (ablation) | Progressive vision-aided training · FFHQ 1k | FID: 15.48 -> 11.43 | Helped |
| Fine-tune the ViT CLIP feature extractor instead of freezing it (ablation) | Vision-aided discriminator · FFHQ 1k, LSUN Cat 1k, and LSUN Cat full data | FID: FID increased to twice the baseline | Hurt |
| Replace pretrained ViT CLIP weights with random weights (ablation) | Vision-aided discriminator · FFHQ 1k, LSUN Cat 1k, and LSUN Cat full data | FID: 11.63 -> 19.10; 15.49 -> 33.77; 4.61 -> 6.35 | Hurt |
| Use two original-architecture discriminators instead of an ensemble including pretrained features (ablation) | StyleGAN2-ADA · FFHQ 1k, LSUN Cat 1k, and LSUN Cat full data | FID: 11.63 -> 17.59; 15.49 -> 37.01; 4.61 -> — | Hurt |
| Train StyleGAN2-ADA longer with standard adversarial loss rather than adding pretrained features (ablation) | StyleGAN2-ADA · FFHQ 1k, LSUN Cat 1k, and LSUN Cat full data | FID: 11.63 -> 19.07; 15.49 -> 39.36; 4.61 -> 6.52 | Hurt |
| Use a multi-scale rather than single-scale ViT CLIP discriminator head (ablation) | ViT CLIP vision-aided discriminator · LSUN Cat 1k and FFHQ 1k | FID: 17.21 -> 15.49; 13.31 -> 11.63 | Helped |
| Remove differentiable data augmentation (ablation) | Vision-aided GAN training · Limited-sample settings | FID | Hurt |
| Remove label smoothing (ablation) | Vision-aided GAN training · LSUN Cat 1k | FID | Hurt |
Closest papers by meaning
- When, Why, and Which Pretrained GANs Are Useful? (ICLR 2022 · similarity 0.92)
- Instance Selection for GANs (NeurIPS 2020 · similarity 0.89)
- Small-GAN: Speeding up GAN Training using Core-Sets (ICML 2020 · similarity 0.89)
- Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator (ICML 2025 · similarity 0.89)
- Mixed batches and symmetric discriminators for GAN training (ICML 2018 · similarity 0.89)
- Sequential Training of GANs Against GAN-Classifiers Reveals Correlated "Knowledge Gaps" Present Among Independently Trained GAN Instances (CVPR 2023 · similarity 0.89)
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.