Research map / Architectures and efficiency / Pruning and efficient tuning
Don’t Be So Dense: Sparse-to-Sparse GAN Training Without Sacrificing Performance
Shiwei Liu, Yuesong Tian, Tianlong Chen, Li Shen
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 sparse parameter exploration (STU-GAN) to the sparse generator, s_D=70 (ablation) | Static Sparse GAN (BigGAN) · CIFAR-10, s_G=80% | FID: 10.03 -> 8.57 | Improved |
| add sparse parameter exploration with extremely sparse generator and dense discriminator (ablation) | Static Sparse GAN (BigGAN) · CIFAR-10, s_G=95%, s_D=0 | FID: 22.5 to 12.5 | Improved |
| add sparse parameter exploration, s_D=70 (ablation) | Static Sparse GAN (SNGAN) · CIFAR-10, s_G=95% | FID: 23.52 -> 20.67 | Improved |
| add sparse parameter exploration, s_D=50 (ablation) | Static Sparse GAN (BigGAN) · CIFAR-10, s_G=90% | FID: 15.37 -> 10.42 | Improved |
| train sparse from scratch with STU-GAN instead of iterative pruning and retraining (GAN Tickets) | STU-GAN s_D=70 vs GAN Tickets (BigGAN) · CIFAR-10, s_G=90% | FID: 9.87 -> 9.75 | Improved |
| train sparse from scratch with STU-GAN instead of one-shot pruning and fine-tuning | PF (uniform) vs STU-GAN s_D=70 (BigGAN) · CIFAR-10, s_G=95% | FID: 16.32 -> 13.27 | Improved |
| sparse discriminator with 80% sparse generator and 70% sparse discriminator versus dense BigGAN | STU-GAN (BigGAN) · CIFAR-10 | FID: 8.62 -> 8.57 | Improved |
| explore parameters in both G and D instead of only G (s_D=50) (ablation) | STU-GAN · CIFAR-10 | FID | Hurt |
| explore parameters only in the discriminator (ablation) | STU-GAN · CIFAR-10 | FID | Hurt |
| extend training to 5x steps with same exploration frequency (ablation) | STU-GAN s_D=50 (BigGAN) · CIFAR-10, s_G=95% | FID: 13.47 -> 12.92 | Improved |
| extend training to 5x steps with same exploration frequency (ablation) | STU-GAN s_D=50 (SNGAN) · CIFAR-10, s_G=95% | FID: 22.34 -> 18.44 | Improved |
| vary exploration frequency Delta T (ablation) | STU-GAN · CIFAR-10 | FID | Mixed |
Closest papers by meaning
- Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis (CVPR 2025 · similarity 0.91)
- RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 2024 · similarity 0.90)
- A Large-Scale Study on Regularization and Normalization in GANs (ICML 2019 · similarity 0.90)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.89)
- The GAN is dead; long live the GAN! A Modern GAN Baseline (NeurIPS 2024 · similarity 0.89)
- Compressing Image-to-Image Translation GANs Using Local Density Structures on Their Learned Manifold (AAAI 2024 · 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.
