Nonmonotonic Sampling Fails to Improve Diffusion Models
A new study systematically tests whether monotonic noise reduction in diffusion models is necessary. Researchers designed four families of nonmonotonic schedules and tested them on DDPM, EDM, and Flow Matching across 90 configurations on CIFAR-10. None outperformed the monotonic baseline, though penalty magnitude varied. The work challenges a core assumption in generative modeling.
Key facts
- Diffusion models traditionally use monotonically decreasing noise levels.
- The study tested nonmonotonic schedules on DDPM, EDM, and Flow Matching.
- Experiments covered NFE budgets from 10 to 200 function evaluations.
- A 42-cell hyperparameter ablation was performed.
- All 90 configurations showed no improvement over monotonic sampling.
- The penalty magnitude varied across configurations.
- The research was published on arXiv with ID 2605.11773.
- The study used CIFAR-10 as the dataset.
Entities
Institutions
- arXiv