Research map / Image generation / Fast diffusion sampling
MAD: Manifold Attracted Diffusion
Dennis Elbrächter, Giovanni S Alberti, Matteo Santacesaria
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 |
|---|---|---|---|
| replace standard score inference with extended score inference (MAD) | DDPM++ trained on noisy synthetic shapes · synthetic data, Gaussian noise | FID: 319.42 -> 189.51 | Improved |
| replace standard score inference with MAD | DDPM++ trained on noisy CIFAR-10 · CIFAR-10, Gaussian noise | FID: 147.47 -> 77.15 | Improved |
| replace standard score inference with MAD | DDPM++ trained on noisy FFHQ · FFHQ, Gaussian noise | FID: 134.42 -> 74.52 | Improved |
| replace standard score inference with MAD on non-Gaussian corruption | DDPM++ trained on FFHQ with pixel removal · FFHQ, pixel removal | FID: 387.87 -> 104.00 | Improved |
| replace standard score inference with MAD on shot noise | DDPM++ trained on FFHQ with shot noise · FFHQ, shot noise | FID: 257.42 -> 101.12 | Improved |
| replace standard score inference with MAD on pixel removal for synthetic data | DDPM++ trained on synthetic shapes · synthetic data, pixel removal | FID: 234.40 -> 163.14 | Improved |
| train diffusion model on noisy data and use standard score inference | DDPM++ · all noisy datasets | FID vs clean data: 320.56 -> 319.42 (synthetic Gaussian training data vs standard generation) | No clear effect |
| increase parameter b in extended score inference (ablation) | MAD on pretrained FFHQ and AFHQv2 networks · noiseless FFHQ/AFHQv2 | background and fine features | Mixed |
| use extended score only in the last half of inference steps (ablation) | MAD · inference cost | compute overhead: 100% -> 50% additional cost | Improved |
| compute exact derivative of score network via automatic differentiation instead of finite difference (ablation) | MAD · numerical experiments | sample quality and cost | No clear effect |
| apply MAD to Cryo-EM EMPIAR-11618 with a network trained only on noisy data | DDPM++ · EMPIAR-11618 | qualitative particle shapes | Improved |
Closest papers by meaning
- Diffusion Models Encode the Intrinsic Dimension of Data Manifolds (ICML 2024 · similarity 0.92)
- Improving the Euclidean Diffusion Generation of Manifold Data by Mitigating Score Function Singularity (NeurIPS 2025 · similarity 0.92)
- Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models (NeurIPS 2024 · similarity 0.91)
- Diffusion Models and the Manifold Hypothesis: Log-Domain Smoothing is Geometry Adaptive (NeurIPS 2025 · similarity 0.90)
- Manifolds, Random Matrices and Spectral Gaps: The geometric phases of generative diffusion (ICLR 2025 · similarity 0.90)
- CW Complex Hypothesis for Image Data (ICML 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.
