score_sde_ve_gen

→ScoreSDEForImageGeneration

score_sde_ve_gen(pretrained: bool = False, overrides: object = {})

source edit

The VE model with its objective and the three samplers.

Model Size

Parameters

pretrainedbool= False

No weights are published for this family; passing True raises.

**overridesobject= {}

Optional ScoreSDEConfig field overrides.

Returns

ScoreSDEForImageGeneration

Trainable and samplable.

Notes

Reference: Song et al., ICLR 2021 (arXiv:2011.13456).

The VE prior is wide — $\mathcal{N}(0, \sigma_{\max}^2 I)$ — so a sampler that starts from a standard normal will not work here, which is why the prior belongs to the SDE rather than the sampler.

Examples

>>> from lucid.models import score_sde_ve_gen
>>> model = score_sde_ve_gen().eval()
>>> model.sde.__class__.__name__
'VESDE'

Used by 1

lucid.models.generative.score_sde