Research map / Image generation / GANs
Train Sparsely, Generate Densely: Memory-Efficient Unsupervised Training of High-Resolution Temporal GAN
Masaki Saito, Shunta Saito, Masanori Koyama, Sosuke Kobayashi
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 frame subsampling layers (s_t=2) (ablation) | multi-scale temporal GAN · UCF101 (4 GPUs) | IS: 13.29 -> 23.87 | Improved |
| increase s_t from 2 to 4 (ablation) | multi-scale temporal GAN · UCF101 (4 GPUs) | IS: 23.87 -> 26.60 | Improved |
| add subsampling layers under a single GPU (ablation) | multi-scale temporal GAN · UCF101 (1 GPU) | IS: 1.43 -> 20.61 | Improved |
| add subsampling layers at batch size 32 (ablation) | multi-scale temporal GAN · UCF101 (batch size 32) | IS: 14.44 -> 23.87 | Improved |
| add subsampling layers at batch size 4 (ablation) | multi-scale temporal GAN · UCF101 (batch size 4) | IS: 13.29 -> 14.64 | Improved |
| add class conditioning | multi-scale temporal GAN · UCF101 (batch 64, s_t=2) | IS: 24.39 -> 54.93 | Improved |
| add class conditioning | multi-scale temporal GAN · UCF101 (batch 64, s_t=4) | FID: 3334 -> 3352 | No clear effect |
| enable first subsampling layer S1 (ablation) | multi-scale temporal GAN · UCF101 (batch 16, s_t=2) | IS: 12.79 -> 19.99 | Improved |
| enable S2 after S1 (ablation) | multi-scale temporal GAN · UCF101 (batch 16, s_t=2) | IS: 19.99 -> 21.38 | Improved |
| enable S3 after S1 and S2 (ablation) | multi-scale temporal GAN · UCF101 (batch 16, s_t=2) | IS: 21.38 -> 22.81 | Improved |
| enable S3 in addition to S1 (s_t=4) (ablation) | multi-scale temporal GAN · UCF101 (batch 16, s_t=4) | IS: 20.95 -> 19.37 | Hurt |
| apply a single subsampling layer to baseline single-discriminator model (s_t=2) (ablation) | 3D discriminator baseline · UCF101 (batch 16) | IS: 11.10 -> 4.30 | Hurt |
Closest papers by meaning
- Memorization Precedes Generation: Learning Unsupervised GANs with Memory Networks (ICLR 2018 · similarity 0.91)
- Temporal Generative Adversarial Nets With Singular Value Clipping (ICCV 2017 · similarity 0.90)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.89)
- Don’t Be So Dense: Sparse-to-Sparse GAN Training Without Sacrificing Performance (IJCV 2023 · similarity 0.89)
- Towards Faster and Stabilized GAN Training for High-fidelity Few-shot Image Synthesis (ICLR 2021 · similarity 0.89)
- RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 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.
