mmagic.models.editors.disco_diffusion.secondary_model

Module Contents

Classes

ConvBlock	Convolution Block.
SkipBlock	Skip block wrapper. Wrapping main block and skip block and concat their
FourierFeatures	Fourier features mapping MLP.
SecondaryDiffusionImageNet2	A smaller secondary diffusion model trained by Katherine Crowson to

Functions

append_dims(x, n)	Append dims.
expand_to_planes(x, shape)	Expand tensor to planes.
alpha_sigma_to_t(alpha, sigma)	convert alpha&sigma to timestep.
t_to_alpha_sigma(t)	convert timestep to alpha and sigma.

mmagic.models.editors.disco_diffusion.secondary_model.append_dims(x, n)

Append dims.

mmagic.models.editors.disco_diffusion.secondary_model.expand_to_planes(x, shape)

Expand tensor to planes.

mmagic.models.editors.disco_diffusion.secondary_model.alpha_sigma_to_t(alpha, sigma)

convert alpha&sigma to timestep.

mmagic.models.editors.disco_diffusion.secondary_model.t_to_alpha_sigma(t)

convert timestep to alpha and sigma.

class mmagic.models.editors.disco_diffusion.secondary_model.ConvBlock(c_in, c_out)

Bases: torch.nn.Sequential

Convolution Block.

Parameters

• c_in (int) – Input channels.

• c_out (int) – Output channels.

class mmagic.models.editors.disco_diffusion.secondary_model.SkipBlock(main, skip=None)

Bases: torch.nn.Module

Skip block wrapper. Wrapping main block and skip block and concat their outputs together.

Parameters

• main (list) – A list of main modules.

• skip (nn.Module) – Skip Module. If not given, set to nn.Identity(). Defaults to None.

forward(input)

Forward function.

class mmagic.models.editors.disco_diffusion.secondary_model.FourierFeatures(in_features, out_features, std=1.0)

Bases: torch.nn.Module

Fourier features mapping MLP.

Parameters

• in_features (int) – Input channels.

• out_features (int) – Output channels.

• std (float) – Standard deviation. Defaults to 1..

forward(input)

Forward function.

class mmagic.models.editors.disco_diffusion.secondary_model.SecondaryDiffusionImageNet2

Bases: torch.nn.Module

A smaller secondary diffusion model trained by Katherine Crowson to remove noise from intermediate timesteps to prepare them for CLIP.

Ref: https://twitter.com/rivershavewings/status/1462859669454536711 # noqa

forward(input, t)

Forward function.