Research map / Image generation / Fast diffusion sampling
Understanding Flatness in Generative Models: Its Role and Benefits
Taehwan Lee, Kyeongkook Seo, Jaejun Yoo, Sung Whan Yoon
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 SAM (ablation) | ADM · CIFAR-10 (20 steps) | FID: 34.47 -> 9.01 | Improved |
| add SAM (ablation) | ADM · CIFAR-10 (100 steps) | FID: 8.80 -> 3.83 | Improved |
| add SAM (ablation) | ADM · LSUN Tower (20 steps) | FID: 36.65 -> 16.02 | Improved |
| add SAM (ablation) | ADM · FFHQ (20 steps) | FID: 30.81 -> 11.59 | Improved |
| add EMA (ablation) | ADM · CIFAR-10 (20 steps) | FID: 34.47 -> 10.63 | Improved |
| add SAM on top of EMA (ablation) | ADM+EMA · CIFAR-10 (20 steps) | FID: 10.63 -> 7.00 | Improved |
| add SAM on top of EMA (ablation) | ADM+EMA · FFHQ (100 steps) | FID: 6.19 -> 5.04 | Hurt |
| add input perturbation (IP) (ablation) | ADM · FFHQ (100 steps) | FID: 7.53 -> 13.55 | Hurt |
| add input perturbation (IP) (ablation) | ADM · CIFAR-10 (20 steps) | FID: 34.47 -> 20.11 | Improved |
| add SAM to measure flatness (LPF) (ablation) | ADM · CIFAR-10 flatness | LPF: 0.097 -> 0.063 | Improved |
| add EMA or SWA to ADM for flatness (LPF) (ablation) | ADM · CIFAR-10 flatness | LPF: 0.097 -> 0.099 | No clear effect |
| add SAM to ADM under 8-bit quantization (ablation) | ADM · CIFAR-10 (8-bit quantized, 20 steps) | FID: 48.02 -> 8.94 | Improved |
Closest papers by meaning
- Sharpness-Aware Minimization Enhances Feature Quality via Balanced Learning (ICLR 2024 · similarity 0.89)
- Sharpness-Aware Minimization Efficiently Selects Flatter Minima Late In Training (ICLR 2025 · similarity 0.89)
- Unveiling m-Sharpness Through the Structure of Stochastic Gradient Noise (NeurIPS 2025 · similarity 0.88)
- On Inference Stability for Diffusion Models (AAAI 2024 · similarity 0.88)
- Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.87)
- Tilted Sharpness-Aware Minimization (ICML 2025 · similarity 0.87)
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.
