Sampler settings

The fine controls a sampler exposes: how much fresh noise it adds back, when, and which internal solver it runs.

Most samplers take no settings beyond steps and schedule. The ones that do expose these:
  • eta and noise scale, on the samplers that add noise back as they go: the ancestral and SDE families. eta decides whether and how much of that noise comes back, and noise scale multiplies it.
  • churn, on Euler and Heun. It is a controlled amount of noise stirred in across a window of the run, set by the two churn bounds.
  • the solver order, on multistep solvers such as DEIS and IPNDM. A few of those also offer a choice of internal correction: midpoint or Heun, phi_1 or phi_2.
Every field is optional. Left blank, the worker uses the sampler's own default, and every other render in this library was made that way.The examples move one setting at a time with the rest left at their defaults. The useful room around each default is narrow: a little either way is a visible edit, and a lot of it is grain or a flat picture.
The same picture rendered at 0
Eta 0
The same picture rendered at 2.5
Eta 2.5
Sampling · Interactive explorer Eta eta controls how much fresh noise an SDE sampler injects as it steps.
The same picture rendered at 0
Noise scale 0
The same picture rendered at 1.5
Noise scale 1.5
Sampling · Interactive explorer Noise scale Noise scale controls how strongly an ancestral or SDE sampler adds fresh noise back after each step.
The same picture rendered at 0
Churn 0
The same picture rendered at 2
Churn 2
Sampling · Interactive explorer Churn Churn deliberately adds noise between steps, then lets the sampler remove it again.
The same picture rendered at midpoint
Solver type midpoint
The same picture rendered at heun
Solver type heun
Sampling · Interactive explorer Solver type The dpmpp_2m_sde family can apply either of two second-order corrections between steps.