Sampler

The algorithm that turns the starting noise into an image.

Speed depends on the choice Sets the character of the image; quality is decided elsewhere

Samplers differ in character and in how quickly they settle on a picture. There is no best one, and k_euler and k_dpmpp_2m are safe starting points.What matters for the rest of the panel is how your sampler answers

, and samplers fall into two groups.A sampler that settles walks steadily towards one image. Raise the steps and you refine the picture you were already getting, and the gain tails off once the picture stops changing.A sampler that re-rolls adds a small amount of fresh noise back at every pass. Raise the steps and you get a different picture. The extra passes do not make the same picture cleaner, so a step count that suited a sampler that settles may look unfinished here.Both behaviours are correct, and the difference decides whether it is worth paying for more steps. The names tell you which you have: an _a suffix marks an ancestral sampler and sde marks a stochastic one, and both re-roll.

The samplers that re-roll are k_euler_a, k_dpm_2_a, k_dpmpp_2s_a and k_dpmpp_sde.
The samplers that settle are k_euler, k_dpm_2, k_dpmpp_2m, k_heun, k_lms, DDIM, dpmsolver and lcm.

Some models name the samplers they were tuned for, and the panel marks those as suggested.

Sampling · Interactive explorer Sampler