#

Latent Diffusion Models

Latent diffusion models use an auto-encoder to map between image space and latent space. The diffusion model works on the latent space, which makes it a lot easier to train. It is based on paper High-Resolution Image Synthesis with Latent Diffusion Models.

They use a pre-trained auto-encoder and train the diffusion U-Net on the latent space of the pre-trained auto-encoder.

For a simpler diffusion implementation refer to our DDPM implementation. We use same notations for $α_{t}$ , $β_{t}$ schedules, etc.

24from typing import List
25
26import torch
27import torch.nn as nn
28import torch.nn.functional
29
30from labml_nn.diffusion.stable_diffusion.model.autoencoder import Autoencoder
31from labml_nn.diffusion.stable_diffusion.model.clip_embedder import CLIPTextEmbedder
32from labml_nn.diffusion.stable_diffusion.model.unet import UNetModel

#

This is an empty wrapper class around the U-Net. We keep this to have the same model structure as CompVis/stable-diffusion so that we do not have to map the checkpoint weights explicitly.

35class DiffusionWrapper(nn.Module):

#

43    def __init__(self, diffusion_model: UNetModel):
44        super().__init__()
45        self.diffusion_model = diffusion_model

#

47    def forward(self, x: torch.Tensor, time_steps: torch.Tensor, context: torch.Tensor):
48        return self.diffusion_model(x, time_steps, context)

#

Latent diffusion model

This contains following components:

51class LatentDiffusion(nn.Module):

#

61    model: DiffusionWrapper
62    first_stage_model: Autoencoder
63    cond_stage_model: CLIPTextEmbedder

#

unet_model is the U-Net that predicts noise $ϵ_{c ond} (x_{t}, c)$ , in latent space
autoencoder is the AutoEncoder
clip_embedder is the CLIP embeddings generator
latent_scaling_factor is the scaling factor for the latent space. The encodings of the autoencoder are scaled by this before feeding into the U-Net.
n_steps is the number of diffusion steps $T$ .
linear_start is the start of the $β$ schedule.
linear_end is the end of the $β$ schedule.

65    def __init__(self,
66                 unet_model: UNetModel,
67                 autoencoder: Autoencoder,
68                 clip_embedder: CLIPTextEmbedder,
69                 latent_scaling_factor: float,
70                 n_steps: int,
71                 linear_start: float,
72                 linear_end: float,
73                 ):

#

85        super().__init__()

#

Wrap the U-Net to keep the same model structure as CompVis/stable-diffusion.

88        self.model = DiffusionWrapper(unet_model)

#

Auto-encoder and scaling factor

90        self.first_stage_model = autoencoder
91        self.latent_scaling_factor = latent_scaling_factor

#

CLIP embeddings generator

93        self.cond_stage_model = clip_embedder

#

Number of steps $T$

96        self.n_steps = n_steps

#

$β$ schedule

99        beta = torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_steps, dtype=torch.float64) ** 2
100        self.beta = nn.Parameter(beta.to(torch.float32), requires_grad=False)

#

$α_{t} = 1 - β_{t}$

102        alpha = 1. - beta

#

$\overset{α_{t}}{ˉ} = \prod_{s = 1}^{t} α_{s}$

104        alpha_bar = torch.cumprod(alpha, dim=0)
105        self.alpha_bar = nn.Parameter(alpha_bar.to(torch.float32), requires_grad=False)

#

Get model device

107    @property
108    def device(self):

#

112        return next(iter(self.model.parameters())).device

#

Get CLIP embeddings for a list of text prompts

114    def get_text_conditioning(self, prompts: List[str]):

#

118        return self.cond_stage_model(prompts)

#

Get scaled latent space representation of the image

The encoder output is a distribution. We sample from that and multiply by the scaling factor.

120    def autoencoder_encode(self, image: torch.Tensor):

#

127        return self.latent_scaling_factor * self.first_stage_model.encode(image).sample()

#

Get image from the latent representation

We scale down by the scaling factor and then decode.

129    def autoencoder_decode(self, z: torch.Tensor):

#

135        return self.first_stage_model.decode(z / self.latent_scaling_factor)

#

Predict noise

Predict noise given the latent representation $x_{t}$ , time step $t$ , and the conditioning context $c$ .

$ϵ_{c ond} (x_{t}, c)$

137    def forward(self, x: torch.Tensor, t: torch.Tensor, context: torch.Tensor):

#

146        return self.model(x, t, context)