Best Scheduler + Sampler Combos

Sampler and scheduler choices change how the image is denoised. Most combinations can run, but only a smaller set is consistently useful for quality, speed, or specific model families.

Use these combinations as reliable starting points. If an image looks noisy, flat, or over-smoothed, try one of the pairings below before changing the prompt.

Quick Picks

GoalSamplerSchedulerNotes
Sharp general outputeulernormalFast, predictable, and a good baseline.
Balanced high qualitydpmpp_2mkarrasStrong default for SDXL-style work.
Realistic depth and shadingdpmpp_2m_sdekarrasGood for portraits, materials, and lighting.
Complex SDXL scenesdpmpp_3m_sdelinear_quadraticBetter for rich scenes and gradients.
Fast low-step generationlcmsgm_uniformBest when speed matters more than maximum detail.
Clean progressive outputdeissimpleUseful for stable text-to-image results.
Controlled diffusion stepsipndmddim_uniformGood compromise when you want tighter noise control.
Sequential or turbo workflowsres_multistepkarrasUseful for Z-Image style or sequential inference workflows.
Anime or 3D-looking renderser_sdeexponentialCan produce smooth depth and stylized volume.
High-resolution adaptive outputuni_pckl_optimalBest when the model supports it well.

Model Starting Points

Model familyRecommended starter
Stable Diffusion 1.5ddim + normal, or euler + normal for sharper output
Flux1 Devddim + sgm_uniform
Flux2 Kleineuler + flux2; try euler + beta for an alternate scheduler
Animaer_sde + simple; try er_sde + exponential for more dimensional renders
Qwen Imageeuler + simple
Z-Imageres_multistep + simple; try res_multistep + karras for a different motion/detail balance
SDXLdpmpp_2m + karras, or dpmpp_3m_sde + linear_quadratic for complex scenes
Pony Diffusiondpmpp_2m + karras, or euler + normal for a simpler baseline

Detailed Pairing List

SamplerBest schedulerWhen to use it
eulernormalFast, sharp, reliable baseline.
euler_cfg_ppkarrasCFG-heavy workflows that need better detail retention.
euler_ancestralexponentialSofter images, dreamy lighting, gradual transitions.
dpmpp_2mkarrasBalanced quality and stability.
dpmpp_2m_sdekarrasRealism, depth, and detailed lighting.
dpmpp_3m_sdelinear_quadraticComplex SDXL scenes and smooth gradients.
heunpp2karrasIntricate prompts and cleaner token-weight transitions.
lcmsgm_uniformFast low-step generation.
uni_pckl_optimalAdaptive high-resolution workflows.
deissimpleClean, progressive text-to-image output.
ipndmddim_uniformNoise-controlled diffusion steps.
res_multistepkarrasSequential inference and animation-like workflows.
er_sdeexponentialSmooth depth, stylized renders, and SDXL variants.

Avoid These Unless You Are Testing

  • Avoid lcm with exponential, kl_optimal, or linear_quadratic; it is designed for fast, low-step generation.
  • Avoid uni_pc with simple or ddim_uniform if the result looks flat.
  • Avoid GPU/SDE samplers with high-noise schedulers unless you intentionally want unstable or experimental output.
  • Avoid CFG-specific samplers unless your workflow is built for that conditioning style.

References