Sampler Parameters

The algorithm used to progressively denoise an image during generation. Different samplers produce varying results in quality, style, and speed.

Samplers determine how the AI removes noise at each step. Choosing the right sampler affects image quality, generation speed, and artistic style.

Fast Samplers (10-20 steps)

  • Euler: Simple, fast, good baseline
  • Euler a: Euler with ancestral sampling, more creative
  • DPM++ 2M: Excellent quality/speed balance
  • DPM++ SDE: More detailed, slightly slower

Quality Samplers (20-40 steps)

  • DPM++ 2M Karras: Smooth, detailed results
  • DPM++ 3M SDE: High detail, slower
  • UniPC: Fast convergence, good quality

Specialty Samplers

  • DDIM: Deterministic, good for interpolation
  • LMS: Classic sampler, smooth results
  • Heun: Higher quality, 2x slower

Ancestral vs Non-Ancestral

  • Ancestral (a): Adds randomness, never fully converges
  • Non-ancestral: Deterministic, converges to final image

Karras vs Linear

Schedulers that modify noise reduction:

  • Karras: Front-loads denoising, often better quality
  • Linear: Even denoising across steps

Recommendations

Use CaseSamplerSteps
Quick previewEuler15
BalancedDPM++ 2M Karras25
High qualityDPM++ SDE Karras30-40
ConsistencyDDIM25
Related terms: Sampling Steps Scheduler CFG Scale