Source
Martin Garcia et al. (2024; WACV 2025), LIT-687, measuring the rule Lin et al. (2023; WACV 2024), LIT-689, stated: “sample steps should always include the last timestep t = T”. Lin et al. named the leading, linspace and trailing spacings. They credit the trailing discretization itself to DPM-Solver (LIT-076), which this record has not checked.
What it costs, measured before anyone knew: Marigold’s authors report about 34% NYUv2 AbsRel at one step (LIT-691, Fig. 7) on weights that reach 5.7 at one step once the spacing is trailing.
What to do
Choose inference timesteps so the first network call gets the timestep that
matches the pure-noise input it is given: [T, …], not […, 1]. With T = 1000 and k steps, trailing gives [1000, 1000 − T/k, …] and leading gives
[…, 1], missing T entirely. In diffusers this is
timestep_spacing="trailing".
Why
The network is trained on pairs of noise level and timestep. Leading spacing
at small k breaks that pairing at the first step, where the input is pure
noise and the timestep says otherwise. For one step, “the model receives a
timestep encoding that indicates an almost perfect depth map whereas the
actual input is pure noise”. The error is largest when every step counts
and when the first prediction is expected to be nearly final, which is the
case for single-step inference and strongly conditioned models.
Conditions
- It matters at few steps and fades toward T. “In the limit of k → T inference steps, both strategies converge.” That describes the timestep list. On a zero-terminal-SNR model, Lin et al.’s one same-seed image shows leading still changing the composition at 25 steps, because the single pure-noise timestep is exactly the one leading skips. That is one image, so it is a hedge, not a claim.
- For text-to-image generation LIT-687 calls the gain “slight”. That is its characterization of Lin et al. Lin never measures spacing alone. Nothing in either paper tests unconditional generation. For image-conditional prediction it is the difference between output and noise.
- It is not zero-terminal-SNR rescaling. Lin et al. propose both, and they are separate fixes. This practice is the spacing alone. SOTA-263’s zero-terminal-SNR discussion is about the other one.
Known implementations
- diffusers exposes the setting, and its DDIM default is still
leading. Marigold’s v1-1 configuration opts in (adoption,DP-005).