Understanding Samplers in Stable Diffusion

The long list of samplers available in GUIs like ComfyUI and Automatic1111 can be intimidating. However, the real-world difference in output is impactful, but not nearly as impactful as a good model or prompt.

Note: Sampler performance varies significantly between models and even between fine-tunes of the same base model. The recommendations below are starting points—always experiment with your specific model to find optimal settings.

How Sampling Works

At the highest level, when you generate an image, the diffusion model starts with random noise and gradually removes it in steps until a clear image forms—a process called sampling.

Different sampling methods, or samplers, remove noise in their own ways, affecting speed and image quality.

Popular Samplers Comparison

Sampler	Steps	Speed	Quality	Notes
Euler	20-30	Fast	Good	Simple, deterministic
DPM++ 2M Karras	20-30	Fast	Great	Balanced speed and quality
DPM++ SDE Karras	10-15	Slow	Best	Prioritizes quality over speed
UniPC	20-30	Fast	Good	Newer sampler, good all-around choice

Recommendations by Model Family

SD 1.5

The most forgiving model family—nearly all samplers work well.

• Recommended: DPM++ 2M Karras @ 20-30 steps
• Alternative: Euler a @ 20-30 steps for faster iteration
• CFG Scale: 7-8

SDXL

Requires slightly different tuning than SD 1.5.

• Recommended: DPM++ 2M Karras @ 20-30 steps
• Alternative: Euler @ 25-35 steps
• CFG Scale: 5-7 (lower than SD 1.5)
• Note: If using base + refiner workflow, the refiner typically needs fewer steps (10-15)

Flux (Dev/Schnell)

Flux uses flow matching rather than traditional diffusion, so sampler behavior differs significantly.

• Flux Dev: Euler @ 20-28 steps
• Flux Schnell: Euler @ 4 steps (optimized for speed)
• CFG Scale: 1-3.5 (much lower than SD models)
• Note: Karras schedulers have minimal effect on Flux; stick with default/normal scheduler

SD3 / SD3.5

Uses rectified flow like Flux.

• Recommended: Euler @ 20-30 steps
• Alternative: DPM++ 2M @ 25-35 steps
• CFG Scale: 4-7
• Note: Often benefits from slightly higher step counts than SDXL

Z-Image Turbo

• Recommended: Res Multistep @ 9 steps
• CFG Scale: 1

Steps and Quality

More steps usually mean better quality but slower generation, while fewer steps are faster but may produce less detailed images. However, there's a point of diminishing returns—going beyond 40-50 steps rarely improves output and can sometimes degrade it.

Ancestral vs Non-Ancestral Samplers

Ancestral samplers (marked with "a" like Euler a, DPM2 a) add random noise during each sampling step. This can create interesting variations but makes the final image non-deterministic—you'll get different results even with the same seed. Non-ancestral samplers are more predictable and consistent, which is better for reproducible workflows.

Quick Start Recommendations

Goal	Sampler	Steps
Fast iteration/testing	Euler	15-20
Balanced performance	DPM++ 2M Karras	20-30
Highest quality	DPM++ SDE Karras	25-40
Flux models	Euler	20-28 (Dev) / 4 (Schnell)

Experiment to find what works best for your specific model and use case! By understanding samplers and steps, you can fine-tune your results and create the images you want more efficiently.