Research map / Image generation / Fast diffusion sampling
Smoothing the Score Function to Enhance Generalization in Diffusion Models
Xinyu Zhou, Jiawei Zhang, Stephen J. Wright
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 |
|---|---|---|---|
| remove noise conditioning (Noise Unconditioning) (ablation) | VE-SDE · CIFAR-10 (SDE PC 1K NFE) | FID(G, Ttr): 6.49 -> 7.33 | Hurt |
| remove noise conditioning (Noise Unconditioning) (ablation) | VE-SDE · CelebA 64x64 (SDE PC 1K NFE) | FID(G, Ttr): 7.25 -> 7.07 | Improved |
| remove noise conditioning (Noise Unconditioning) (ablation) | VE-SDE · CelebA 64x64 (ODE PC 1K NFE) | FID(G, Ttr): 7.65 -> 7.71 | No clear effect |
| use feature-space KNN instead of pixel-space KNN for temperature smoothing (ablation) | Temperature Smoothing (Ti=7/σi, K=100) · CIFAR-10 (SDE PC 1K NFE) | FID(G, Ttr): 50.81 -> 7.96 | Improved |
| use feature-space KNN instead of pixel-space KNN for temperature smoothing (ablation) | Temperature Smoothing (Ti=10/σi, K=100) · CelebA 64x64 (SDE PC 1K NFE) | FID(G, Ttr): 61.97 -> 8.40 | Improved |
| increase temperature and K in temperature smoothing with pixel-space KNN (ablation) | Temperature Smoothing · CIFAR-10 (SDE PC 1K NFE) | FID(G, Ttr): 8.32 -> 50.81 | Hurt |
| add temperature smoothing (Ti=5/σi, K=30, feature KNN) to unconditioning (ablation) | Noise Unconditioning · CelebA 64x64 (SDE PC 1K NFE) | FID(G, Ttr): 7.07 -> 8.78 | Hurt |
| increase temperature in score weights (ablation) | empirical score function · Cat-Caracal (sampling expansiveness) | expansion ratio γex | Improved |
| remove noise conditioning in neural network score (ablation) | NN score function · Cat-Caracal (1,200 images) | memorization (qualitative) | Improved |
Closest papers by meaning
- Generalization through variance: how noise shapes inductive biases in diffusion models (ICLR 2025 · similarity 0.91)
- Understanding Generalizability of Diffusion Models Requires Rethinking the Hidden Gaussian Structure (NeurIPS 2024 · similarity 0.91)
- Generalization Dynamics of Linear Diffusion Models (NeurIPS 2026 · similarity 0.90)
- On the Edge of Memorization in Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Neural Network-Based Score Estimation in Diffusion Models: Optimization and Generalization (ICLR 2024 · similarity 0.89)
- Understanding diffusion models requires rethinking (again) generalization (NeurIPS 2026 · 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.
