Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Exploring the Optimal Choice for Generative Processes in Diffusion Models: Ordinary vs Stochastic Differential Equations
Yu Cao, Jingrun Chen, Yixin Luo, Xiang ZHOU
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 |
|---|---|---|---|
| raise h^2/g^2 from 0 (ODE) to 2 at 3000 steps (ablation) | pre-trained VP model · CIFAR-10, 3000 steps | FID: 7.27 -> 6.69 | Improved |
| raise h^2/g^2 from 0 (ODE) to 1 at 1000 steps (ablation) | pre-trained VP model · CIFAR-10, 1000 steps | FID: 6.89 -> 6.70 | Improved |
| raise h^2/g^2 from 0 (ODE) to 2 at 100 steps (ablation) | pre-trained VP model · CIFAR-10, 100 steps | FID: 22.43 -> 52.77 | Hurt |
| raise h^2/g^2 from 0 to 4 at 500 steps (ablation) | pre-trained VP model · CIFAR-10, 500 steps | FID: 7.20 -> 20.47 | Hurt |
| raise h^2/g^2 from 0 to 4 at 4000 steps (ablation) | pre-trained VP model · CIFAR-10, 4000 steps | FID: 7.33 -> 7.01 | No clear effect |
| increase discretization steps from 1000 to 4000 for ODE (ablation) | ODE sampler (h=0) · CIFAR-10 | FID: 6.89 -> 7.33 | Hurt |
| increase discretization steps from 100 to 2000 for h^2/g^2 = 4 (ablation) | SDE sampler · CIFAR-10 | FID: 92.83 -> 7.20 | Improved |
| increase diffusion coefficient h in the generative process (ablation) | Swiss roll score model · Swiss roll | Wasserstein distance | Improved |
| increase number of discretization steps for ODE (h=0) (ablation) | Swiss roll score model · Swiss roll | Wasserstein distance | No clear effect |
| increase diffusion coefficient h (ablation) | 2D 4-mode Gaussian mixture with specified score error · 2D Gaussian mixture | KL divergence | Improved |
Closest papers by meaning
- On the Trajectory Regularity of ODE-based Diffusion Sampling (ICML 2024 · similarity 0.91)
- Towards Non-Asymptotic Convergence for Diffusion-Based Generative Models (ICLR 2024 · similarity 0.90)
- Diffusion models for Gaussian distributions: Exact solutions and Wasserstein errors (ICML 2025 · similarity 0.90)
- Evaluating the design space of diffusion-based generative models (NeurIPS 2024 · similarity 0.90)
- Adaptive Stochastic Coefficients for Accelerating Diffusion Sampling (NeurIPS 2025 · similarity 0.89)
- A solvable model of learning generative diffusion: theory and insights (NeurIPS 2025 · 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.
