What is Sampling in AI Image Generation? Complete Guide to Samplers and Schedulers

Every time you generate an AI image, a sampler is doing the heavy lifting. It's the algorithm that actually removes noise step-by-step to create your image. Understanding samplers unlocks better results and faster generations. Here's everything you need to know.

Quick Answer: Sampling is the iterative process of removing noise from random static to reveal an image. Samplers are the algorithms that decide how to remove noise at each step. Schedulers control how much noise to remove per step. Together, they determine image quality, generation speed, and artistic characteristics. For most users, DPM++ 2M with Karras scheduler offers the best balance of speed and quality.

What Sampling Actually Does

The Noise Removal Process

AI image generation starts with pure random noise, like TV static. The model's job is to progressively remove that noise while following your prompt, eventually revealing an image.

Each "step" in generation:

1. Model looks at current noisy image
2. Predicts what noise should be removed
3. Sampler removes calculated amount of noise
4. Process repeats until complete

The sampler determines HOW noise gets removed. Different samplers take different mathematical approaches, producing different results.

Why Samplers Matter

Two samplers with identical prompts and seeds produce noticeably different images:

• Composition may shift
• Colors appear different
• Details vary in sharpness
• Artistic style changes subtly

Choosing the right sampler can mean the difference between okay results and exceptional ones.

Common Samplers Explained

Euler (The Classic)

The simplest sampler, named after mathematician Leonhard Euler.

Characteristics:

• Clean, straightforward results
• Good baseline quality
• Moderate speed
• Predictable behavior

Best for: Learning, testing, when you want predictable results

Steps recommended: 20-30

Euler Ancestral (Euler a)

Euler with added randomness at each step.

Characteristics:

• More creative, varied results
• Higher variability between generations
• Same seed produces slightly different images
• More artistic, less predictable

Best for: Creative exploration, artistic variation

Steps recommended: 20-40

Note: "Ancestral" samplers add randomness, making them less reproducible.

DPM++ 2M

Currently the most popular sampler. "DPM" stands for Diffusion Probabilistic Models.

Characteristics:

• Excellent quality-to-speed ratio
• Converges quickly (needs fewer steps)
• Clean, detailed outputs
• Reliable across different prompts

Best for: General use, production workflows, most content types

Steps recommended: 15-25 (converges fast)

DPM++ 2M SDE

DPM++ 2M with stochastic differential equations, adding controlled randomness.

Characteristics:

• Slightly more detailed than standard DPM++ 2M
• More variation between generations
• Can produce more interesting textures
• Slightly slower

Best for: When you want more variation or texture detail

Steps recommended: 20-30

DPM++ 2S a

Another DPM++ variant using second-order solving with ancestral sampling.

Characteristics:

• Good detail retention
• Ancestral (adds randomness)
• Can produce unique results
• Less predictable than 2M

Best for: Artistic work, creative exploration

Steps recommended: 20-35

DDIM (Denoising Diffusion Implicit Models)

Deterministic sampler that produces consistent results.

Characteristics:

• Completely deterministic (same seed = exact same image)
• Fast at low step counts
• Can feel "flat" compared to others
• Good for reproducibility

Best for: When exact reproducibility matters

Steps recommended: 20-50

UniPC (Unified Predictor-Corrector)

Newer sampler designed for efficiency.

Characteristics:

• Very fast convergence
• Good quality at low step counts
• Works well with 10-15 steps
• Consistent results

Best for: Speed-focused workflows, quick iterations

Steps recommended: 10-20

Heun

Higher-order solver producing detailed results.

Characteristics:

• Excellent detail
• Slower than Euler/DPM++
• Uses more VRAM
• Sharp, crisp outputs

Best for: When maximum detail matters more than speed

Steps recommended: 20-40 (but takes twice as long)

Understanding Schedulers

Schedulers control the RATE of noise removal across steps. Same sampler with different schedulers produces different results.

Normal Scheduler

Linear noise schedule, removes consistent amount per step.

Characteristics:

• Straightforward approach
• Consistent step-to-step changes
• Baseline behavior

Best for: When you want standard behavior

Karras Scheduler

Named after researcher Tero Karras. Removes more noise early, less late.

Characteristics:

• Front-loads noise removal
• Produces sharper results
• Better detail preservation
• Most popular scheduler choice

Best for: Most use cases, especially with DPM++ samplers

Exponential Scheduler

Exponential curve for noise removal.

Characteristics:

• Aggressive early denoising
• Smooth final steps
• Can produce cleaner backgrounds

Best for: Experimentation, specific aesthetic goals

SGM Uniform

Uniform scheduler from Score-based Generative Models.

Characteristics:

• Even distribution across steps
• Predictable behavior
• Works well with specific samplers

Best for: Technical workflows requiring consistency

Sampler + Scheduler Combinations

The combination matters. Here are proven pairings:

DPM++ 2M + Karras (Recommended Default)

The gold standard combination for most work.

• Fast convergence
• Excellent quality
• Reliable results
• 15-25 steps optimal

Euler + Normal

Classic, predictable combination.

• Baseline behavior
• Good for learning
• 20-30 steps typical

DPM++ SDE + Karras

For when you want variation and detail.

• More texture
• Slight randomness
• 20-30 steps optimal

UniPC + Normal

Speed-focused combination.

• Fastest good results
• 10-15 steps works
• Good for iteration

Choosing the Right Sampler

By Content Type

Portraits:

• Recommended: DPM++ 2M Karras or Euler
• Avoid: Highly ancestral samplers (unpredictable faces)

Landscapes:

• Recommended: DPM++ 2M SDE Karras (good textures)
• Alternative: Euler a for artistic variation

Anime/Illustration:

• Recommended: DPM++ 2M Karras
• Alternative: Euler a for more stylized results

Photorealistic:

• Recommended: DPM++ 2M Karras
• Alternative: Heun for maximum detail (slower)

Abstract/Artistic:

• Recommended: Euler a or DPM++ 2S a
• More randomness suits creative work

By Priority

Speed Priority:

• UniPC (10-15 steps)
• DPM++ 2M Karras (15-20 steps)

Quality Priority:

• Heun (40+ steps)
• DPM++ 2M Karras (25-30 steps)

Balanced:

• DPM++ 2M Karras (20 steps)
• Default choice for most work

Creative Variation:

• Euler a
• DPM++ 2S a
• Any ancestral sampler

Step Count Guide

More steps = more noise removal iterations = better quality (diminishing returns).

Step Count Recommendations

Diminishing Returns

Most samplers converge around 20-30 steps. After that, quality improvements are minimal while time increases linearly.

Exception: Some samplers (DDIM, Euler) benefit from higher steps more than others (DPM++, UniPC).

Practical Workflow Tips

Testing Samplers

When trying a new sampler:

1. Use same seed across tests
2. Start with 20 steps
3. Compare results side-by-side
4. Adjust steps based on convergence

Production Settings

For reliable production work:

• Sampler: DPM++ 2M
• Scheduler: Karras
• Steps: 20
• CFG Scale: 7 (adjust as needed)

This combination works reliably across most content types.

Speed vs Quality Trade-offs

Need maximum speed?

• UniPC, 10 steps
• Accept slightly lower quality

Need maximum quality?

• DPM++ 2M Karras, 30 steps
• Or Heun, 40 steps (much slower)

Need balance?

• DPM++ 2M Karras, 20 steps
• Best overall workflow

Sampler Settings in ComfyUI

In ComfyUI's KSampler node:

• sampler_name: Choose your sampler (euler, dpmpp_2m, etc.)
• scheduler: Choose scheduler (karras, normal, etc.)
• steps: Number of denoising iterations
• cfg: Classifier-free guidance (prompt adherence)
• denoise: For img2img workflows

For ComfyUI workflow optimization, see our workflow organization guide.

Common Problems and Solutions

Blurry Results

Cause: Too few steps or wrong sampler

Solution: Increase steps to 25-30, use DPM++ 2M Karras

Overcooked/Burnt Look

Cause: Too many steps or CFG too high

Solution: Reduce steps, lower CFG scale

Inconsistent Results

Cause: Using ancestral sampler

Solution: Switch to deterministic sampler (DPM++ 2M, DDIM) for consistency

Slow Generation

Cause: High steps or slow sampler

Solution: Use UniPC or DPM++ 2M with fewer steps

Weird Artifacts

Cause: Sampler-model mismatch or extreme settings

Solution: Try different sampler, reset to defaults

Advanced Sampler Concepts

Deterministic vs Ancestral

Deterministic samplers:

• Same seed = exact same result
• Examples: DPM++ 2M, DDIM, UniPC
• Best for reproducibility

Ancestral samplers (marked with 'a'):

• Add randomness at each step
• Same seed = similar but not identical
• Examples: Euler a, DPM++ 2S a
• Best for creative variation

Solver Orders

Some samplers mention "order":

• First-order: Simple, fast (Euler)
• Second-order: More accurate, slower (Heun, some DPM++ variants)
• Higher order generally means better quality at cost of speed

SDE vs ODE

ODE (Ordinary Differential Equations):

• Deterministic
• Consistent results

SDE (Stochastic Differential Equations):

• Add controlled noise
• More variation
• Can produce more interesting details

Frequently Asked Questions

Which sampler is "best"?

DPM++ 2M Karras is best for most use cases. No single sampler is best for everything.

Does sampler affect generation speed?

Yes. Heun is 2x slower than Euler. UniPC is fastest. Most DPM++ variants are moderately fast.

Can I change samplers mid-workflow?

Not directly, but you can use different samplers for img2img refinement passes.

Do samplers work with all models?

Yes, samplers are model-agnostic. Results vary by model though.

What's the difference between sampler and scheduler?

Sampler: HOW noise is removed. Scheduler: HOW MUCH noise per step.

Why do ancestral samplers give different results with same seed?

They intentionally add randomness for creative variation. It's a feature, not a bug.

How do I know when a sampler has converged?

Run same prompt at different step counts. When results stop improving noticeably, you've converged.

Should I always use Karras scheduler?

For DPM++ samplers, usually yes. For others, test both Normal and Karras.

Do video generation models use the same samplers?

Similar concepts apply, but video models often have specialized sampling implementations.

What sampler does Midjourney use?

Midjourney uses proprietary sampling methods not publicly disclosed.

Wrapping Up

Samplers and schedulers are fundamental to AI image generation. Understanding them helps you:

• Generate faster with fewer steps
• Get more consistent results
• Achieve specific artistic effects
• Troubleshoot quality issues

Key recommendations:

• Default to DPM++ 2M Karras at 20 steps
• Use Euler a for creative exploration
• Use UniPC when speed matters most
• Increase steps only when quality requires it

For hands-on generation without worrying about sampler details, Apatero.com handles optimization automatically. For deeper technical control, ComfyUI exposes all sampler options.

Quick Reference: Sampler Comparison

Start with DPM++ 2M Karras. Experiment from there based on your specific needs and the results you're getting.