PROMPT GUIDE

Sampler vs Scheduler: The Difference and How to Choose (Euler, DPM++, Karras)

In Stable Diffusion settings, "sampler" and "scheduler" both decide how a picture progresses while it is being made. The names are similar and are often written side by side, as in Euler a or DPM++ 2M Karras, which makes this the place where beginners get confused most. This guide sums up the difference in their roles and how to choose when you cannot decide.

Last updated: 2026-10-22 approx. 3 min read

日本語版

The short answer

The sampler is the procedure that turns noise into a picture, and the scheduler is how the noise is divided up across that procedure. When in doubt, try DPM++ 2M or Euler a as the sampler and Karras or Simple (Normal) as the scheduler, and adjust the step count from around 20 to 30. Which look you prefer changes with each model, so the shortcut is to line them up with the same settings and compare.

What is the difference between a sampler and a scheduler?

The sampler is the method that turns noise into a picture, and the scheduler is how much noise is reduced at each step.

An image generation AI makes a picture by removing noise that looks like static, little by little, over several steps. The sampler (sampling method) decides how the noise is removed in a single pass, and the scheduler (schedule type) decides when to remove a lot and when to remove only a little. In older screens they were combined into one name such as DPM++ 2M Karras, but newer screens may split them into two boxes. How they are displayed depends on the tool and version, so check the wording on your screen.

Item	What it decides	Examples
Sampler	How noise is removed (the way it is calculated)	Euler a, DPM++ 2M, DDIM and so on
Scheduler	How the removal is divided across steps	Karras, Exponential, Simple (Normal) and so on
Steps	How many passes the removal is divided into	Try from around 20 to 30

Sampler: DPM++ 2M / Schedule type: Karras / Steps: 24 / CFG scale: 6

How do you choose among the main samplers?

If you cannot decide, start with either DPM++ 2M or Euler a and switch by the look you like.

Samplers come in two kinds: the "ancestral" kind, with an a at the end of the name, and the ordinary kind without it. Ancestral samplers add noise back in along the way, so the picture keeps changing even when you raise the step count and never converges. In exchange, they can produce soft-looking pictures and unexpected compositions. Ordinary samplers get closer to the same picture as you raise the step count, so they suit finishing one picture in fine detail.

Sampler	Trait	Good for
Euler a	Ancestral. Light, and the picture changes when you change the steps	Brainstorming ideas, making many pictures casually
Euler	Ordinary. Simple and stable	Finishing the same picture bit by bit
DPM++ 2M	Ordinary. Tends to come together even in few steps	Basic finishing, reproducing the same picture
DPM++ SDE	Adds noise. Details change easily	When you want to try textures (tends to take time)
DDIM, UniPC	Easy to use with few steps	When you prioritize speed

Rinka's tips

Sampler names differ a little between tools and versions. The ones you can use appear on your screen, so first look for the names in the table above
The model description page sometimes lists a recommended sampler. Try those settings first

Anime girl smiling on a riverside path, a reference image for comparing samplers — 見比べる絵（1）

The same place with a slightly different result, showing how samplers and steps change a picture — 見比べる絵（1）

Illustration usage rules

How do you choose a scheduler such as Karras or Simple?

If you cannot decide, use Karras or leave the tool's default (Simple, Normal or Automatic), and change it only when the picture looks wrong.

A scheduler is an allocation table that changes how noise is reduced between the first and second half of the steps. Karras, for example, gives more time to the fine finishing in the second half, so it is said to bring out details nicely. Which one fits depends on how it works with the model, so there is no "right answer."

Steps for deciding the sampler and scheduler

First make one picture with the tool's default (or the settings in the model description)
Fix the seed and make two or three pictures changing only the sampler
Keep the sampler that gave the picture you like best, then change the scheduler and compare again
Compare steps at 20, 25 and 30 and use the point where the picture stops changing
Write down the combination you chose along with the name of the model

For how to fix the seed and match conditions, see What Is a Seed? Fixed, Random, and Reproducing the Same Image →

For how to line up settings with an XYZ plot to compare them, see How to Use the XYZ Plot (Compare Settings Side by Side) →

What if the picture does not turn out clean? (Common failures and fixes)

Check in this order: too few steps, where to use an ancestral sampler, and the balance with the CFG scale.

Common failure	Cause	Fix
The picture changes every time you raise the steps	Ancestral samplers such as Euler a do not converge	Use an ordinary sampler such as DPM++ 2M when you want to finish the same picture
The picture is grainy and details collapse	Too few steps, or the sampler does not fit	Set the steps to 24 to 30 and try DPM++ 2M with Karras
Colors are harsh and lines are hard	The CFG scale is too high	Lower the CFG scale, then compare samplers
It takes too long	A heavy sampler or too many steps	Use Euler a or DPM++ 2M and reduce the steps
It becomes a different picture even with supposedly the same settings	One of the sampler or scheduler differs	Check the two settings in the generation info and match them

Sampler: Euler a / Schedule type: Automatic / Steps: 20 / CFG scale: 5 (for brainstorming)

For how to decide together with the CFG scale and steps, see What Is CFG Scale? How to Choose Steps and Samplers →

For the settings of high-resolution finishing, see Hires.fix Settings (Upscale Ratio, Denoising Strength, Upscaler) →

To find the cause from the symptoms of a broken picture, see AI Art Troubleshooting Guide (Look Up Fixes by Symptom) →

FAQ

Frequently Asked Questions

Which sampler do you recommend?

If you cannot decide, start with DPM++ 2M or Euler a. DPM++ 2M suits refining the same picture, and Euler a suits casually trying many different pictures. If the model description has recommended settings, using those first is the surest.

What does the a in Euler a mean?

It is the first letter of ancestral, a method that adds noise back in along the way. The picture does not converge even when you raise the step count, and keeps changing.

Must I always use Karras?

It is not required. It is said to bring out details nicely, but with some models the default scheduler fits better. Line them up with the same seed and decide by comparing.

I cannot find a scheduler box.

Depending on the version, Karras and similar names are included in the sampler name. In that case you choose by one name such as DPM++ 2M Karras. Displays differ between versions and tools, so check the explanation for the screen you use.

Are more steps always better?

No. Beyond a certain point the change becomes small and only the time grows. Try from around 20 to 30 steps, and the point where the picture stops changing is enough.