Research map / Image generation / Fast diffusion sampling
Gaussian Mixture Solvers for Diffusion Models
Hanzhong Allan Guo, Cheng Lu, Fan Bao, Tianyu Pang, Shuicheng YAN, Chao Du, Chongxuan Li
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 Gaussian reverse kernel with 2-component Gaussian mixture (GMS) vs SN-DDPM | GMS · CIFAR10 (LS), K=10 | FID: 21.87 -> 17.43 | Improved |
| GMS vs SN-DDPM at 10 steps, cosine schedule | GMS · CIFAR10 (CS), K=10 | FID: 16.33 -> 13.80 | Improved |
| GMS vs best prior SDE solver at 10 steps | GMS · CIFAR10 | FID: 4.44 | Improved |
| GMS vs SN-DDPM at 1000 steps, linear schedule | GMS · CIFAR10 (LS), K=1000 | FID: 3.71 -> 2.76 | Improved |
| GMS vs SN-DDPM at 1000 steps, cosine schedule | GMS · CIFAR10 (CS), K=1000 | FID: 4.08 -> 4.23 | Hurt |
| GMS vs SN-DDPM at 25 steps | GMS · ImageNet 64x64, K=25 | FID: 27.58 -> 26.50 | Improved |
| GMS vs SN-DDPM at 4000 steps | GMS · ImageNet 64x64, K=4000 | FID: 16.22 -> 15.79 | Improved |
| use analytic variance (SN-DDPM) instead of fixed beta_t in DDPM at 10 steps | DDPM variants · CIFAR10 (LS), K=10 | FID: 233.41 -> 21.87 | Improved |
| DDPM beta_t vs beta-tilde_t at 1000 steps | DDPM variants · CIFAR10 (LS), K=1000 | FID: 5.11 -> 3.04 | Improved |
| increase SDEdit start time t0 from 0.3T to larger | DDPM variants · SDEdit stroke synthesis | realism | Hurt |
| GMS vs SN-DDPM in SDEdit at same t0 and steps | GMS · SDEdit stroke synthesis | faithfulness / realism | Mixed |
| increase sampling steps for SDE-based solver vs ODE-based DDIM | DDPM variants · SDEdit stroke synthesis | realism | Mixed |
Closest papers by meaning
- Diffusion models for Gaussian distributions: Exact solutions and Wasserstein errors (ICML 2025 · similarity 0.90)
- End-to-end Learning of Gaussian Mixture Priors for Diffusion Sampler (ICLR 2025 · similarity 0.89)
- TADA: Improved Diffusion Sampling with Training-free Augmented DynAmics (NeurIPS 2025 · similarity 0.88)
- MaRS: A Fast Sampler for Mean Reverting Diffusion based on ODE and SDE Solvers (ICLR 2025 · similarity 0.88)
- Diffusion Bridge Implicit Models (ICLR 2025 · similarity 0.88)
- Adaptive Stochastic Coefficients for Accelerating Diffusion Sampling (NeurIPS 2025 · similarity 0.88)
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.
