data

StableDiffusionOutput

extendsModelOutput

StableDiffusionOutput(noise_pred: Tensor, latent: Tensor, loss: Tensor | None = None)

source edit

What StableDiffusionModel returns.

Attributes

noise_predTensor

(B, latent_channels, h, w) —

\epsilon_\theta

latentTensor

The noised latent the prediction was made from.

lossTensor or None

The denoising objective, present only when a target was given.

Examples

>>> import lucid
>>> from lucid.models.generative.stable_diffusion._model import (
...     StableDiffusionOutput,
... )
>>> out = StableDiffusionOutput(
...     noise_pred=lucid.zeros(1, 4, 32, 32),
...     latent=lucid.zeros(1, 4, 32, 32),
... )
>>> out.noise_pred.shape
(1, 4, 32, 32)
>>> out.loss is None
True
Both live in latent space, not pixels — decoding is a separate step,
and doing it per training step would cost more than the training.

Used by 1

lucid.models.generative.stable_diffusion

Constructors

dunder

init

→None

__init__(noise_pred: Tensor, latent: Tensor, loss: Tensor | None = None)

source edit

Initialise the model. See the class docstring for parameters.

>>> import lucid >>> from lucid.models.generative.stable_diffusion._model import ( ... StableDiffusionOutput, ... ) >>> out = StableDiffusionOutput( ... noise_pred=lucid.zeros(1, 4, 32, 32), ... latent=lucid.zeros(1, 4, 32, 32), ... ) >>> out.noise_pred.shape (1, 4, 32, 32) >>> out.loss is None True Both live in latent space, not pixels — decoding is a separate step, and doing it per training step would cost more than the training.