Sampler and sampler names
The node that turns noise into a picture step by step, and what the names in its sampler_name list mean in practice.
In plain words
Generation starts from pure noise. At each step the model looks at the noisy latent and predicts what noise is in it; the sampler removes part of that noise and moves on. After the last step a clean latent remains, and VAE Decode turns it into a picture.
The sampler node is where everything meets: the model, the positive and negative conditioning, the starting latent and the settings. The sampling algorithm (sampler_name) is the rule for how each step moves; the scheduler decides how big each step is.
Where you meet it in a graph
- KSampler: inputs
model,positive,negative,latent_image; fieldsseed,control_after_generate,steps,cfg,sampler_name,scheduler,denoise; outputLATENT. - KSampler (Advanced): adds
add_noise,start_at_step,end_at_step,return_with_leftover_noise, to split one generation between two samplers (base and refiner, or the two Wan 2.2 models). - SamplerCustom and SamplerCustomAdvanced: assembled from parts:
NOISE(RandomNoise),GUIDER(CFGGuider, BasicGuider),SAMPLER(KSamplerSelect),SIGMAS(BasicScheduler and others). Flux and many newer templates use this form. - Packs have their own samplers (for example WanVideo Sampler).
Sampler names in practice
| Name | What people use it for |
|---|---|
euler | Simple and dependable. The default in many Flux, Wan, Qwen-Image and Z-Image templates. A good first choice. |
euler_ancestral (_a) | Adds fresh noise every step: more variety, softer look; the picture keeps changing as you change steps. |
dpmpp_2m | Popular with SDXL and SD 1.5, usually with the karras scheduler; sharp at 20 to 30 steps. |
dpmpp_2m_sde, dpmpp_sde, dpmpp_3m_sde | Stochastic: more fine detail and variation; slower per step. |
uni_pc, uni_pc_bh2 | Converge in few steps; used in some video graphs. |
ddim | The classic; mostly for old graphs. |
lcm | Only for LCM models and LCM LoRAs, at 4 to 8 steps and cfg about 1 to 2. |
res_multistep, er_sde, gradient_estimation and others | Newer methods; use them when a template or model page names them. |
Names ending in _ancestral or _sde add noise during sampling, so the same seed gives slightly different results with different step counts. Names ending in _gpu do the same work with noise made on the graphics card.
Typical mistakes
- Changing sampler and scheduler at random: on a well-trained model the difference is small next to the prompt and the seed. Start from the model's page.
- An ancestral sampler when comparing settings: results jump. Use
eulerfor comparisons. lcmon an ordinary model: a blurry, unfinished picture.
How the app helps
The node description of every sampler explains its inputs. The lessons The sampler's settings and Choosing a sampler and a scheduler go through its fields; in KSampler (Advanced) graphs the video lessons check that the two samplers hand over at the same step. In Properties, a field with a built-in note has a button What this field does next to its name; it opens the note under the field: what the field does and what to put in it.