SDXL Samplers Explained — DPM++, Euler, and What Actually Matters

The one rule

Use DPM++ 2M Karras at 25–30 steps. Move on.

That's 90 % of the answer for 90 % of prompts. Everything below is when to deviate.

The three samplers worth knowing

DPM++ 2M Karras — the sensible default. Fast, high quality, works well at 25–30 steps. Slightly deterministic (same prompt + same seed = same image every time).

Euler a (Euler ancestral) — stochastic. Same prompt + same seed will give slightly different images. Slower to converge but sometimes more creative. Use for exploration.

DPM++ SDE Karras — the "when quality matters most" pick. Slower than DPM++ 2M, marginally sharper output.

Step counts

• Under 15 steps: noticeable degradation
• 20–30 steps: sweet spot for most samplers
• 30–50 steps: diminishing returns
• 50+ steps: you're wasting compute

The rest of the sampler zoo

DDIM, Heun, LMS, UniPC, DPM++ 3M SDE... they exist. They rarely matter. If someone's telling you a specific niche sampler is the secret to their images, the actual secret is usually their prompt.

CFG scale (the other dial)

Sampler and step count are half the answer. CFG scale (7 is the SDXL default) controls how strictly the model follows your prompt. Higher = closer to your prompt, more likely to look "burned." Lower = more creative freedom, more drift.

• CFG 5–6: soft, painterly
• CFG 7–8: default sweet spot
• CFG 9–11: sharp, sometimes over-cooked
• CFG 12+: usually broken

Imagoat's Generate page uses these defaults automatically. Advanced controls (step count, CFG, sampler) are on the roadmap for the Pro plan.