What is the difference between ancestral and non-ancestral samplers?

Ancestral samplers (names ending in 'a', like Euler a) inject noise at each step, so the image keeps changing and never fully converges. Non-ancestral samplers (Euler, DPM++ 2M, DDIM) converge to a stable result as steps increase, which is better for reproducibility and upscaling.

What is the safest all-round sampler?

DPM++ 2M Karras is the common default — it gives sharp, coherent results in 20–30 steps and converges cleanly, which makes it reliable for both first passes and hires fix.

Why does the same seed look different on another sampler?

Each sampler integrates the denoising path differently, so the same seed and prompt produce a different image per sampler. Lock the sampler along with the seed when you need reproducibility.

Is anything sent to a server?

No. This is a static reference that runs in your browser. Nothing is uploaded, generated or logged.

What is the SD Sampler Selector Guide?

Free Stable Diffusion sampler guide. Tell it whether you want speed, quality or consistency and it recommends a sampler (Euler a, DPM++ 2M Karras, DDIM and more) with a comparison table and step ranges — all in your browser. It runs free in your browser on Gera Tools, with nothing uploaded.

SD Sampler Selector Guide

Name: SD Sampler Selector Guide
Creator: Gera Tools
License: https://creativecommons.org/licenses/by/4.0/

Stable Diffusion sampler selector guide

The sampler decides how Stable Diffusion turns noise into your image, and the right choice changes both speed and look. Tell this guide whether you want speed, quality or consistency, and it recommends a sampler with a sensible step range — backed by a full comparison table.

How samplers differ

A sampler is the numerical solver that walks the image from pure noise toward the denoised result. They fall into two broad families:

Ancestral / stochastic (Euler a, DPM++ 2S a) inject fresh noise each step. They produce lively, varied output but never fully converge — adding steps keeps changing the image.
Convergent (Euler, DPM++ 2M Karras, DDIM, UniPC) settle toward a stable result. These are best for reproducibility, hires fix and upscaling, because more steps refine rather than redraw.

The Karras label refers to a noise schedule that spends more steps where they matter most, which is why DPM++ 2M Karras reaches clean results in relatively few steps.

Sampler comparison at a glance

Sampler	Speed	Quality	Converges?	Best for
Euler a	Fast	Good	No	Exploration, varied outputs
Euler	Fast	Good	Yes	Reproducibility, simple baseline
DPM++ 2M Karras	Fast	Excellent	Yes	Default all-round choice
DPM++ 2S a Karras	Medium	Excellent	No	High detail, creative runs
DPM++ 3M SDE	Slow	Very high	Yes	Maximum quality
DDIM	Fast	Good	Yes	Reproducibility, inpainting
UniPC	Fast	Excellent	Yes	Few-step quality

Step suggestions: most convergent samplers reach a clean result at 20–30 steps. Ancestral samplers benefit from more steps (25–40) but keep changing.

The Karras schedule explained

The noise schedule controls how much denoising is done at each step. Standard schedules distribute steps evenly, spending as much time on easy late denoising as on the critical early structure-forming steps. The Karras schedule front-loads more of the budget into the higher-noise early steps where the image’s structure is established, which is why it reaches good results in fewer total steps than an unscheduled sampler. Any sampler with “Karras” in the name uses this schedule; without it, the same solver uses a less efficient default.

When ancestral samplers are useful

Because ancestral samplers inject noise at each step, they produce genuinely different images even when you fix the seed. This makes them good for exploration — generating a batch to see a range of interpretations before committing to a specific look. They are poor choices for hires fix and upscaling, which work by starting partway through the denoising process from a lower-resolution image; a sampler that never converges will change the image unpredictably at that stage rather than refining it.

If you generate a result you love with Euler a and want to reproduce it exactly, note that the output is not fully reproducible even with the same seed — small differences in floating-point arithmetic between runs can produce visible changes. Switch to a convergent sampler to lock the result.

Tips for choosing

Need it fast? Euler a or DPM++ 2M Karras look good by ~20 steps.
Need the best detail? DPM++ 2M Karras or DPM++ 3M SDE at 25–35 steps.
Need it reproducible? Pick a convergent sampler (DDIM, Euler) and lock the seed.
Test, don’t trust. Generate the same prompt across three samplers at a fixed seed and pick the look you prefer — that beats any chart.