Samplers decide the algorithmic path from noise to image. Different samplers vary in convergence, detail and style tendency. This article focuses on the four that matter in production.
Four production samplers
| Sampler | Behavior | Recommended steps | Best for |
|---|---|---|---|
| Euler a | Fast, stylized, non-converging | 20–30 | Illustration, anime |
| DPM++ 2M Karras | Detail-rich, converging | 25–35 | Realistic portrait, product |
| UniPC | Fewest steps to acceptable | 10–20 | Quick drafting |
| DDIM | Deterministic, reproducible | 30–50 | Reproducibility-critical work |
Converging vs non-converging
"Converging" samplers (DPM++ 2M) stabilize and approach a fixed solution as steps increase. "Non-converging" samplers (Euler a, Ancestral) keep introducing noise — every extra step keeps changing the image. Converging samplers are easier to control in production.
Quick picks
- Realistic portrait: DPM++ 2M Karras, 30 steps
- Anime / illustration: Euler a, 25 steps
- Quick draft: UniPC, 15 steps
- Reproducibility: DDIM, 30 steps