Research map / Probabilistic and scientific modeling / Variational and generative sampling
Manifold-Aware Perturbations for Constrained Generative Modeling
Katherine Keegan, Lars Ruthotto
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 |
|---|---|---|---|
| perturb in normal directions instead of isotropic noise (ablation) | DDPM with projection · Plane | TVD: 0.2817 -> 0.2015 | Improved |
| perturb in normal directions instead of isotropic noise (ablation) | DDPM with projection · Sphere | TVD: 0.3097 -> 0.2341 | Improved |
| train on p_sigma instead of p0 then project | Glow · Sphere | TVD: 0.3000 -> 0.2925 | Improved |
| train on p_sigma instead of p0 then project | RealNVP · Sphere | TVD: 0.5969 -> 0.2638 | Improved |
| train on p_sigma instead of p0 then project | Glow · Plane | TVD: 0.5962 -> 0.3187 | Improved |
| train on p_sigma instead of p0 then project | DDPM · Plane | TVD: 0.1893 -> 0.2015 | No clear effect |
| train on p_sigma instead of p0 then project | DDPM · Stanford Bunny mesh | TVD: 0.3458 -> 0.3471 | No clear effect |
| train on p_sigma instead of p0 then project | DDPM · MNIST total flux | class JSD: 0.0869 -> 0.0730 | Improved |
| train on p_sigma instead of p0 then project | DDPM · MNIST total flux | FID: 7.379 × 10−6 -> 1.512×10−5 | Hurt |
| train on p_sigma instead of p0 | DDPM · Protein backbone fragments | MMD: 3.35 × 10−4 -> 3.37 × 10−4 | No clear effect |
| increase sigma toward the manifold reach (ablation) | p_sigma · Sphere (reach 1.0) | coverage, JSD, TVD | Hurt |
| use p_sigma instead of Projected Diffusion Models | DDPM · MNIST total flux | sampling time: 78.1216 -> 39.0945 | Improved |
Closest papers by meaning
- When and how can inexact generative models still sample from the data manifold? (NeurIPS 2025 · similarity 0.88)
- Rectifying Categorical Flows on Statistical Manifolds for One-Step Generation (NeurIPS 2026 · similarity 0.88)
- Distance Marching for Generative Modeling (NeurIPS 2026 · similarity 0.88)
- A Likelihood Based Approach to Distribution Regression Using Conditional Deep Generative Models (ICML 2025 · similarity 0.88)
- Manifolds, Random Matrices and Spectral Gaps: The geometric phases of generative diffusion (ICLR 2025 · similarity 0.87)
- Conditional Diffusion Models are Minimax-Optimal and Manifold-Adaptive for Conditional Distribution Estimation (ICLR 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.
