Research map / Image generation / GANs
Large Scale GAN Training for High Fidelity Natural Image Synthesis
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 |
|---|---|---|---|
| Increase batch size from 256 to 2048 | SA-GAN on ImageNet 128×128 · ImageNet 128×128 | IS: 52.52 -> 76.85 | Helped |
| Increase channel width from 64 to 96 (ablation) | SA-GAN with batch 2048 on ImageNet · ImageNet 128×128 | IS: 76.85 -> 92.98 | Helped |
| Add shared class embeddings (ablation) | BigGAN · ImageNet 128×128 | IS: 92.98 -> 94.94 | Helped |
| Add skip-z connections to multiple generator layers (ablation) | BigGAN with shared embeddings · ImageNet 128×128 | IS: 94.94 -> 98.76 | Helped |
| Add orthogonal regularization to the generator (ablation) | BigGAN with shared embeddings and skip-z · ImageNet 128×128 | IS: 98.76 -> 99.31 | Helped |
| Use four-times-deeper BigGAN-deep architecture (ablation) | BigGAN · ImageNet at 128×128, 256×256, and 512×512 | FID: 8.7±.6 -> 5.7 ± .3 | Helped |
| Sample trained generator inputs from a truncated normal | BigGAN · ImageNet | FID | Mixed |
| Use R1 gradient penalty with strength 10 (ablation) | BigGAN discriminator · ImageNet | IS: 45% reduction in IS | Hurt |
| Reduce R1 penalty strength to 1 (ablation) | BigGAN discriminator · ImageNet | IS: 20% | Hurt |
| Increase model depth (ablation) | BigGAN-deep · JFT-300M 256×256 | FID: 23.48 -> 18.84 | Helped |
| Use Bernoulli {0, 1} latent instead of standard normal (ablation) | SA-GAN baseline · ImageNet | IS: 8% | Helped |
| Use censored normal latent instead of standard normal (ablation) | SA-GAN baseline · ImageNet | IS: 15-20% | Helped |
Closest papers by meaning
- Conditional Image Synthesis with Auxiliary Classifier GANs (ICML 2017 · similarity 0.93)
- Towards Faster and Stabilized GAN Training for High-fidelity Few-shot Image Synthesis (ICLR 2021 · similarity 0.92)
- Projected GANs Converge Faster (NeurIPS 2021 · similarity 0.92)
- Progressive Growing of GANs for Improved Quality, Stability, and Variation (ICLR 2018 · similarity 0.92)
- Scalable GANs with Transformers (ICML 2026 · similarity 0.92)
- Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions (CVPR 2024 · similarity 0.90)
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.