Research map / Image generation / Fast diffusion sampling
One-Line-of-Code Data Mollification Improves Optimization of Likelihood-based Generative Models
Ba-Hien Tran, Giulio Franzese, Pietro Michiardi, Maurizio Filippone
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 annealed Gaussian data mollification during training | Likelihood-based generative models · CIFAR10 and CelebA 64 | FID: 131.15 -> 121.75; 74.62 -> 64.87; 191.98 -> 155.13; 81.25 -> 79.68; 97.59 -> 70.91; 80.19 -> 72.97 | Improved |
| Add Gaussian mollification | NVAE · CIFAR10 | FID: 53.64 -> 52.26 | Improved |
| Add Gaussian mollification | MAF · UCI test sets | test log-likelihood: -16.32 -> -11.51; -14.87 -> -11.96; -8.27 -> -6.17; -13.03 -> -11.65 | Improved |
| Replace Gaussian mollification with blurring mollification (ablation) | Real-NVP · CIFAR10 and CelebA 64 | FID: 121.75 -> 120.88; 79.68 -> 85.40 | Mixed |
| Replace Gaussian mollification with blurring mollification (ablation) | Glow · CIFAR10 and CelebA 64 | FID: 64.87 -> 66.70; 70.91 -> 74.74 | Hurt |
| Replace Gaussian mollification with blurring mollification (ablation) | VAE · CIFAR10 and CelebA 64 | FID: 155.13 -> 175.40; 72.97 -> 77.29 | Hurt |
| Use cosine instead of sigmoid noise schedule (ablation) | Real-NVP · CIFAR10 | FID: 121.75 -> 118.71 | Improved |
| Use linear instead of sigmoid noise schedule (ablation) | Real-NVP · CIFAR10 | FID: 121.75 -> 123.93 | Hurt |
| Use cosine or linear instead of sigmoid noise schedule (ablation) | Glow · CIFAR10 | FID: 64.87 -> 71.90; 64.87 -> 74.36 | Hurt |
| Use cosine instead of sigmoid noise schedule (ablation) | VAE, Beta-VAE, and IWAE · CIFAR10 | FID: 155.13 -> 154.71; 93.90 -> 92.86; 146.70 -> 146.49 | Improved |
| Use linear instead of sigmoid noise schedule (ablation) | VAE, Beta-VAE, and IWAE · CIFAR10 | FID: 155.13 -> 156.47; 93.90 -> 93.14; 146.70 -> 149.16 | Hurt |
| Gradually reduce mollification rather than use a fixed small amount of noise | Likelihood-based generative models · Optimization | optimization performance | Improved |
Closest papers by meaning
- Feature Likelihood Divergence: Evaluating the Generalization of Generative Models Using Samples (NeurIPS 2023 · similarity 0.88)
- A Variational Perspective on Diffusion-Based Generative Models and Score Matching (NeurIPS 2021 · similarity 0.88)
- Preconditioned Score-Based Generative Models (IJCV 2025 · similarity 0.88)
- Classification Diffusion Models: Revitalizing Density Ratio Estimation (NeurIPS 2024 · similarity 0.86)
- Algorithm- and Data-Dependent Generalization Bounds for Diffusion Models (NeurIPS 2025 · similarity 0.86)
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.86)
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.
