Research map / Image generation / Fast diffusion sampling
A Random Matrix Perspective on the Consistency of Diffusion Models
Binxu Wang, Jacob A Zavatone-Veth, Cengiz Pehlevan
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 |
|---|---|---|---|
| use the linear Gaussian (Wiener filter) predictor to predict generated samples | diffusion models trained on disjoint splits · FFHQ32 cross-split MSE | per-sample Pearson r with cross-split consistency: r=0.244 | Improved |
| train on splits stratified along a principal component so mean/variance mismatch (ablation) | UNet and DiT diffusion models · FFHQ32 counterfactual splits | cross-split sample consistency | Hurt |
| increase training set size n from the memorization regime to n >= 3000 (ablation) | UNet/DiT diffusion · FFHQ64 and others | generalization and cross-split consistency | Improved |
| limited training data (finite n) in linear denoiser (ablation) | linear denoiser · FFHQ spectrum | variance along low-variance modes | Hurt |
| increase dataset size on top versus lower eigenmodes (ablation) | UNet/DiT · cross-split MSE per eigenmode | cross-split MSE | Mixed |
| use DiT instead of UNet | diffusion denoiser · FFHQ64 | cross-split MSE | Improved |
| train much longer (250k vs 50k steps) (ablation) | FFHQ64 UNet · FFHQ64 | closeness to linear predictor and cross-split consistency | Hurt |
| amplify or shrink initial noise in top-50/200 eigenspace (ablation) | UNet FFHQ64 sampling · FFHQ64 | visual sample quality | Mixed |
| manipulate initial noise in bottom-200 eigenspace (ablation) | UNet FFHQ64 sampling · FFHQ64 | visual sample change | No clear effect |
Closest papers by meaning
- The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.92)
- Generalization through variance: how noise shapes inductive biases in diffusion models (ICLR 2025 · similarity 0.91)
- An Analytical Theory of Spectral Bias in the Learning Dynamics of Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models (NeurIPS 2024 · similarity 0.91)
- Consistent Diffusion Meets Tweedie: Training Exact Ambient Diffusion Models with Noisy Data (ICML 2024 · similarity 0.91)
- Understanding Generalizability of Diffusion Models Requires Rethinking the Hidden Gaussian Structure (NeurIPS 2024 · similarity 0.90)
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.
