Skip to content
DreamSprite Logo DreamSprite

Samplers & Scheduling

The KSampler (or similar sampling nodes) is where the actual image is created in ComfyUI. It takes your prompt conditioning, ControlNet guidance, latent noise, and model weights, then iteratively denoises random noise into a coherent image over a series of steps.

Two main parts control this process:

  • Sampler: The algorithm that decides how to take each denoising step (e.g., Euler, DPM++, UniPC).
  • Scheduler (or noise schedule): How the noise level decreases over steps (e.g., normal, Karras, exponential).

Together, they determine speed, quality, detail, and how closely the result follows your prompt and ControlNet.

Popular Samplers & Schedulers in 2026

ComfyUI includes dozens, but these are the most used and recommended:

  • Euler a (ancestral) Fast, creative, good for artistic/experimental gens. Often produces vibrant, slightly noisy results. Pairs well with: Karras scheduler. Steps: 20-40.

  • DPM++ 2M Karras The community favorite for balanced quality/speed. Sharp details, good prompt adherence. Steps: 20-40. Scheduler: Karras (built-in).

  • DPM++ SDE Karras Even sharper and more detailed than 2M, but can be noisier or take longer. Great for realism. Steps: 25-50.

  • UniPC / UniPC BH2 Very fast convergence (good quality in fewer steps). Excellent for high-res or when VRAM is tight. Steps: 15-30.

  • LCM / LCM Sampler (Latent Consistency Models) Ultra-fast (4-8 steps) with distilled models. Great for quick previews or low-step workflows. Requires LCM-specific LoRAs/checkpoints.

Scheduler choices (often automatic with Karras samplers):

  • Karras: Smooth, high-quality noise reduction - most popular.
  • Exponential: Stronger early denoising - good for artistic styles.
  • Normal: Classic, but often outperformed by Karras.

Quick recommendation for beginners:

  • Start with DPM++ 2M Karras (20-30 steps) - reliable, good quality, fast enough.
  • Switch to Euler a for more creative/varied outputs.
  • Use UniPC when you want speed without losing much quality.

Key Settings in KSampler

  • Steps: Number of denoising iterations (20-50 typical).

    • More steps = more detail/refinement (but longer generation time).
    • Fewer steps = faster, sometimes more stylized/noisy.

  • CFG Scale (Classifier-Free Guidance): How strongly the model follows your prompt (vs random noise).

    • 5-9: Balanced (most common).
    • 7-8: Sweet spot for most models.
    • <5: More creative/random.

    • 10: Very strict (can cause artifacts or overcooked looks).

    • Tip: Lower CFG (4-6) when using strong ControlNet - lets the control image dominate.

  • Denoise (for img2img/inpainting): How much to change the input image (0.0-1.0).

    • 0.0: No change (just preview).
    • 0.3-0.5: Gentle refinement (fix small flaws).
    • 0.6-0.8: Big changes while keeping composition.
    • 1.0: Full text-to-image (ignores input).

How Samplers Interact with ControlNet & Prompts

  • Strong ControlNet (e.g., OpenPose at 1.0 strength): Lower CFG (5-7) and start ControlNet early (0.0-0.3) - the pose/edges dominate, sampler just fills in details.
  • Weak/no ControlNet: Higher CFG (8-10) to force prompt adherence.
  • Creative vs precise: Ancestral samplers (Euler a) + low CFG = more variation; deterministic samplers (DPM++ 2M) + high CFG = closer to prompt.
  • Timing with ControlNet: Use early ControlNet (pose/depth) for structure, late (lineart/tile) for polish - sampler steps handle the refinement in between.

Quick Tips

  • Always test at low res (512x768) with 20 steps first - saves time.
  • Seed: Fix seed for reproducibility, change for variation.
  • Batch size: 1 for testing, 4+ when you like the settings.
  • Save good combos: Note sampler, steps, CFG, denoise in your Obsidian notes.

Mastering samplers turns “okay” generations into “wow” ones - it’s often the last 20% that makes the difference.