score_sde_subvp(pretrained: bool = False, overrides: object = {})Construct the sub-VP model — the SDE this paper introduces.
Model Size
Parameters
pretrainedbool= FalseNo weights are published for this family; passing
True raises.**overridesobject= {}Optional
ScoreSDEConfig field overrides.Returns
ScoreSDEModelScore network over the sub-VP SDE.
Notes
Reference: Song et al., ICLR 2021 (arXiv:2011.13456), equation 12.
Its variance is bounded by the VP process at every intermediate step, and the paper's best likelihoods — 2.99 bits per dimension on CIFAR-10 — are reported here.
Examples
>>> from lucid.models import score_sde_subvp
>>> model = score_sde_subvp().eval()
>>> model.config.sde_type
'subvp'