Noise schedules and step placement

A common normalized perturbation representation is x_t=α_t x₀+σ_t ε. The training schedule defines how signal and noise coefficients relate to time, and a model learns predictions under that convention. An inference solver selects a finite sequence of compatible states and updates between them. Using fewer indices from an existing schedule is different from changing the underlying α and σ functions without adapting the model. Input preconditioning and output conversion must match the selected coefficients at every actual evaluation point.

Parameterizations such as noise prediction, clean-sample prediction, or velocity prediction are not interchangeable tensor labels. Converting among them can divide by small schedule coefficients and magnify numerical issues near endpoints. Diffusers v0.35.1 exposes scheduler prediction_type and multiple spacing or sigma options, but supported combinations remain scheduler- and model-specific. Inspect the pinned implementation rather than assuming a configuration name carries the same meaning in another library. Record the realized timestep and sigma arrays in an experiment so the grid is reproducible.

Uniform spacing in one coordinate need not be uniform in another because the mapping is nonlinear. A geometric positive-sigma grid allocates equal ratios rather than equal absolute sigma gaps. More points near a regime can reduce a relevant integration defect while spending fewer evaluations elsewhere, but effectiveness depends on the field, solver, conditioning, and model. The DPM-Solver formulation uses a half-log-SNR change of variables to simplify its integral; this structural argument does not declare every geometric grid optimal for every sampler.

Keep initial noise scaling and the final transition explicit. A list of N positive levels has N−1 intervals between those entries, while an implementation may add a zero-noise endpoint, evaluate an extra denoising step, or use higher-order internal calls. Therefore N list entries are not universally N neural evaluations. Never duplicate an endpoint or append a zero that leads to an undefined logarithm merely to match a desired loop length. Verify supported custom-grid APIs and final-step behavior in the exact scheduler version.

Check monotonicity, endpoints, duplicate mapped indices, finite coefficient conversions, and the schedule’s expected input scaling. Compare a candidate with the baseline using paired initial latents, fixed guidance and model revision, and measured evaluations and wall time. A grid that works for an unguided test may behave differently under strong guidance. Preserve the entire configuration rather than reporting only steps=20; otherwise a later implementation can execute a numerically different process under the same shorthand.

The bisection instrument illustrates why step count alone does not determine numerical resolution. Reset, set both c=2 and U=2, predict bracket half-width 0.0625 after four refinements, then Step four times. Separately Reset, set both c=2 and U=4, predict half-width 0.125 after four refinements, then Step four times. Neither run hits the irrational root exactly. This fixed x²−c bisection has no noise schedule, score model, or diffusion update. Its limited comparison is that the same count of refinement operations can cover different coordinate spans and leave different resolution.