Research map / Image generation / Autoregressive and latent generation
On the Anisotropy of Score-Based Generative Models
Andreas Floros, Seyed-Mohsen Moosavi-Dezfooli, Pier Luigi Dragotti
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 |
|---|---|---|---|
| minimize alignment alpha between data and architecture geometry using W_min (ablation) | iDDPM U-Net · MNIST, CelebA-HQ (56x56), CIFAR-10 (SW2) | Sliced Wasserstein-2 | Improved |
| maximize alignment alpha between data and architecture geometry using W_max (ablation) | iDDPM U-Net · MNIST, CelebA-HQ (56x56), CIFAR-10 (SW2) | Sliced Wasserstein-2 | Hurt |
| align rank-one Gaussian data with small-eigenvalue eigenvectors of the average geometry instead of large-eigenvalue ones (ablation) | iDDPM U-Net · Rank-one Gaussian 16x16 data (MSW2) | Max-Sliced Wasserstein-2 | Improved |
| same alignment test on a transformer backbone (ablation) | DiT · Rank-one Gaussian 16x16 data (MSW2) | Max-Sliced Wasserstein-2 | Improved |
| replace nearest resampling with area resampling in the U-Net (ablation) | iDDPM U-Net · Impulse response of symmetric initialization | directional symmetry | Improved |
| model low-frequency data (DCT/DST/Hadamard) versus spatially localized data | iDDPM U-Net · 16x16 rank-one datasets in standard bases (MSW2) | Max-Sliced Wasserstein-2 | Hurt |
| use stochastic gradient descent rather than deterministic GD in anisotropic linear DSM (ablation) | linear DSM with fixed anisotropic output transform · R^5 toy experiment | convergence error | Mixed |
| increase DiT patch size (ablation) | DiT · Rank-one Gaussian 16x16 data | eigenvalue multiplicity of geometry | Hurt |
Closest papers by meaning
- Improved Techniques for Training Score-Based Generative Models (NeurIPS 2020 · similarity 0.89)
- Moderating the Generalization of Score-based Generative Model (ICCV 2025 · similarity 0.88)
- Elucidating the Design Space of Diffusion-Based Generative Models (NeurIPS 2022 · similarity 0.88)
- Beyond Log-Concavity and Score Regularity: Improved Convergence Bounds for Score-Based Generative Models in W2-distance (ICML 2025 · similarity 0.87)
- Algorithm- and Data-Dependent Generalization Bounds for Diffusion Models (NeurIPS 2025 · similarity 0.86)
- A Non-Asymptotic Convergent Analysis for Scored-Based Graph Generative Model via a System of Stochastic Differential Equations (ICML 2025 · 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.
