Sampler

last edited

A sampler is the algorithm that schedules the denoising steps of a diffusion model. Examples include Euler, DPM++, DDIM and UniPC. The model predicts noise, and the sampler decides how to use that prediction to step from one noise level to the next. Sampling can be viewed as solving the corresponding diffusion ordinary differential equations.

Sampler behavior

Different solvers trade speed, sharpness and stability. Ancestral samplers, the "-a" family, inject fresh noise at each step. The image keeps changing and never fully converges. That helps exploration but hurts reproducibility.

The DPM-Solver paper, last revised on 13 October 2022, introduced a fast solver for the diffusion ODE. A separate DPM-Solver++ paper covers the DPM++ variants used in many interfaces.

Schedulers and steps

The scheduler shapes how noise levels are spaced across the step budget. The Karras paper, "Elucidating the Design Space of Diffusion-Based Generative Models," analyzes the design space for sampling, training and preconditioning. Karras-style spacing in interfaces traces back to that work.

Step count is the third setting. Too few steps leave a noisy, unfinished image. Past a model-dependent point, extra steps only add render time.

Distillation and comparison

Distillation cuts step counts sharply. The 2022 progressive-distillation paper reported results with as few as 4 steps. In July 2026, the ICML poster for DMSampler described a few-step distilled sampler for reinforcement learning on diffusion models.

With a fixed seed and prompt, swapping samplers is a clean way to compare their character. Sampler choice also interacts visibly with CFG. See core techniques for related workflows and practical use.

Sources

These primary papers and pages cover diffusion sampling, noise design and fast solvers.