Hands-On Generative AI with

Transformers and Diffusion

Models
With Early Release ebooks, you get books in their earliest form—the
authors’ raw and unedited content as they write—so you can take
advantage of these technologies long before the official release of these
titles.

Pedro Cuenca, Apolinário Passos, Omar

Sanseviero, and Jonathan Whitaker
Hands-On Generative AI with Transformers and Diffusion
Models
by Pedro Cuenca, Apolinário Passos, Omar Sanseviero, and Jonathan
Whitaker
Copyright © 2024 Pedro Cuenca, Apolinário Passos, Omar Sanseviero, and
Jonathan Whitaker. All rights reserved.
Printed in the United States of America.
Published by O’Reilly Media, Inc., 1005 Gravenstein Highway North,
Sebastopol, CA 95472.
O’Reilly books may be purchased for educational, business, or sales
promotional use. Online editions are also available for most titles
(http://oreilly.com). For more information, contact our
corporate/institutional sales department: 800-998-9938 or
[email protected].

Acquisitions Editor: Nicole Butterfield

Development Editor: Jill Leonard

Production Editor: Gregory Hyman

Interior Designer: David Futato

Cover Designer: Karen Montgomery

Illustrator: Kate Dullea

September 2024: First Edition

Revision History for the Early Release

2023-03-16: First Release
See http://oreilly.com/catalog/errata.csp?isbn=9781098149246 for release
details.
The O’Reilly logo is a registered trademark of O’Reilly Media, Inc. Hands-
On Generative AI with Transformers and Diffusion Models, the cover
image, and related trade dress are trademarks of O’Reilly Media, Inc.
The views expressed in this work are those of the authors and do not
represent the publisher’s views. While the publisher and the authors have
used good faith efforts to ensure that the information and instructions
contained in this work are accurate, the publisher and the authors disclaim
all responsibility for errors or omissions, including without limitation
responsibility for damages resulting from the use of or reliance on this
work. Use of the information and instructions contained in this work is at
your own risk. If any code samples or other technology this work contains
or describes is subject to open source licenses or the intellectual property
rights of others, it is your responsibility to ensure that your use thereof
complies with such licenses and/or rights.
978-1-098-14924-6
Chapter 1. Diffusion Models

A NOTE FOR EARLY RELEASE READERS

With Early Release ebooks, you get books in their earliest form—the
authors’ raw and unedited content as they write—so you can take
advantage of these technologies long before the official release of these
titles.
This will be the third chapter of the final book. Please note that the
GitHub repo will be made active later on.
If you have comments about how we might improve the content and/or
examples in this book, or if you notice missing material within this
chapter, please reach out to the editor at [email protected].

In late 2020 a little-known class of models called diffusion models began

causing a stir in the machine-learning world. Researchers figured out how
to use these models to generate synthetic images at higher quality than any
produced by previous techniques. A flurry of papers followed, proposing
improvements and modifications that pushed the quality up even further. By
late 2021 there were models like GLIDE that showcased incredible results
on text-to-image tasks, and a few months later, these models had entered the
mainstream with tools like DALL-E 2 and Stable Diffusion. These models
made it easy for anyone to generate images just by typing in a text
description of what they wanted to see.
In this chapter, we’re going to dig into the details of how these models
work. We’ll outline the key insights that make them so powerful, generate
images with existing models to get a feel for how they work, and then train
our own models to deepen this understanding further. The field is still
rapidly evolving, but the topics covered here should give you a solid
foundation to build on. Chapter 5 will explore more advanced techniques
through the lens of a model called Stable Diffusion, and chapter 6 will
explore applications of these techniques beyond simple image generation.

The Key Insight: Iterative Refinement

So what is it that makes diffusion models so powerful? Previous techniques,
such as VAEs or GANs, generate their final output via a single forward pass
of the model. This means the model must get everything right on the first
try. If it makes a mistake, it can’t go back and fix it. Diffusion models, on
the other hand, generate their output by iterating over many steps. This
‘iterative refinement’ allows the model to correct mistakes made in
previous steps and gradually improve the output. To illustrate this, let’s look
at an example of a diffusion model in action.
We can load a pre-trained model using the Hugging Face diffusers library.
The pipeline can be used to create images directly, but this doesn’t show us
what is going on under the hood:

# Load the pipeline

image_pipe = DDPMPipeline.from_pretrained("google/ddpm-celebahq-
256")
image_pipe.to(device);

# Sample an image
image_pipe().images[0]
We can re-create the sampling process step by step to get a better look at
what is happening as the model generates images. We initialize our sample
x with random noise and then run it through the model for 30 steps. On the
right, you can see the model’s prediction for what the final image will look
like at specific steps - note that the initial predictions are not particularly
good! Instead of jumping right to that final predicted image, we only
modify x by a small amount in the direction of the prediction (shown on the
left). We then feed this new, slightly better x through the model again for
the next step, hopefully resulting in a slightly improved prediction, which
can be used to update x a little more, and so on. With enough steps, the
model can produce some impressively realistic images.

# The random starting point for a batch of 4 images

x = torch.randn(4, 3, 256, 256).to(device)

# Set the number of timesteps lower

image_pipe.scheduler.set_timesteps(num_inference_steps=30)

# Loop through the sampling timesteps

for i, t in enumerate(image_pipe.scheduler.timesteps):

# Get the prediction given the current sample x and the

timestep t
with torch.no_grad():
noise_pred = image_pipe.unet(x, t)["sample"]

# Calculate what the updated sample should look like with the
scheduler
scheduler_output = image_pipe.scheduler.step(noise_pred, t,
x)

# Update x
x = scheduler_output.prev_sample

# Occasionally display both x and the predicted denoised

images
if i % 10 == 0 or i == len(image_pipe.scheduler.timesteps) -
1:
fig, axs = plt.subplots(1, 2, figsize=(12, 5))

grid = torchvision.utils.make_grid(x, nrow=4).permute(1,

2, 0)
axs[0].imshow(grid.cpu().clip(-1, 1) * 0.5 + 0.5)
 axs[0].set_title(f"Current x (step {i})")

pred_x0 = scheduler_output.pred_original_sample
grid = torchvision.utils.make_grid(pred_x0,
nrow=4).permute(1, 2, 0)
axs[1].imshow(grid.cpu().clip(-1, 1) * 0.5 + 0.5)
axs[1].set_title(f"Predicted denoised images (step {i})")
plt.show()

NOTE
Don’t worry if that code looks a bit intimidating - we’ll explain how this all works over
the course of this chapter. For now, just focus on the results.

This core idea of learning how to refine a ‘corrupted’ input gradually can be
applied to a wide range of tasks. In this chapter, we’ll focus on
unconditional image generation - that is, generating images that resemble
the training data, with no additional controls over what these generated
samples look like. Diffusion models have also been applied to audio, video,
text and more. And while most implementations use some variant of the
‘denoising’ approach that we’ll cover here, new approaches utilizing
different types of ‘corruption’ together with iterative refinement are
emerging that may move the field beyond the current focus on denoising
diffusion specifically. Exciting times!

Training a Diffusion Model

In this section, we’re going to train a diffusion model from scratch to gain a
better understanding of how they work. We’ll start by using components
from the Hugging Face diffusers library. As the chapter progresses, we’ll
gradually demystify how each component works. Training a diffusion
model is relatively straightforward compared to other types of generative
models. We repeatedly:

Load in some images from the training data.

Add noise in different amounts. Remember, we want the model to do a
good job estimating how to ‘fix’ (denoise) both extremely noisy
images and images that are close to perfect.
Feed the noisy versions of the inputs into the model.
Evaluate how well the model does at denoising these inputs.
Use this information to update the model weights.

To generate new images with a trained model, we begin with a completely

random input and repeatedly feed it through the model, updating the input
on each iteration by a small amount based on the model prediction. As we’ll
see, there are a number of sampling methods that try to streamline this
process so that we can generate good images with as few steps as possible.
The Data
For this example, we’ll use a dataset of images from the Hugging Face
Hub- specifically, this collection of 1000 butterfly pictures. Later on, in the
projects section, you will see how to use your own data.

dataset = load_dataset("huggan/smithsonian_butterflies_subset",
split="train")

We need to do some preparation before this data can be used to train a

model. Images are typically represented as a grid of ‘pixels’, with color
values between 0 and 255 for each of the three color channels (Red, Green
and Blue). To process these and make them ready for training, we: - Resize
them to a fixed size - (Optional) Add some augmentation by randomly
flipping them horizontally, effectively doubling the size of our dataset -
Convert them to a PyTorch tensor (which represents the color values as
floats between 0 and 1) - Normalize them to have a mean of 0, with values
between -1 and 1
We can do all of this with torchvision.transforms:

image_size = 64

# Define data augmentations

preprocess = transforms.Compose(
[
transforms.Resize((image_size, image_size)), # Resize
transforms.RandomHorizontalFlip(), # Randomly flip (data
augmentation)
transforms.ToTensor(), # Convert to tensor (0, 1)
transforms.Normalize([0.5], [0.5]), # Map to (-1, 1)
]
)

Next, we need to create a dataloader to load the data in batches with these
transforms applied:

batch_size = 32

def transform(examples):
 images = [preprocess(image.convert("RGB")) for image in
examples["image"]]
return {"images": images}

dataset.set_transform(transform)

train_dataloader = torch.utils.data.DataLoader(
dataset, batch_size=batch_size, shuffle=True
)

We can check that this worked by loading a single batch and inspecting the
images.

batch = next(iter(train_dataloader))
print('Shape:', batch['images'].shape,
'\nBounds:', batch['images'].min().item(), 'to',
batch['images'].max().item())
show_images(batch['images'][:8]*0.5 + 0.5) # NB: we map back to
(0, 1) for display
Shape: torch.Size([32, 3, 64, 64])
Bounds: -0.9921568632125854 to 1.0

Adding Noise
How do we gradually corrupt our data? The most common approach is to
add noise to the images. The amount of noise we add is controlled by a
noise schedule. Different papers and approaches tackle this in different
ways, which we’ll explore later in the chapter. For now, let’s see one
common approach in action based on the paper “Denoising diffusion
probabilistic models” by Ho et al. In the diffusers library, adding noise is
handled by something called a scheduler, which takes in a batch of images
and a list of ‘timesteps’ and determines how to create the noisy versions of
those images:

scheduler = DDPMScheduler(num_train_timesteps=1000,
beta_start=0.001, beta_end=0.02)
timesteps = torch.linspace(0, 999, 8).long()

x = batch['images'][:8]
noise = torch.rand_like(x)
noised_x = scheduler.add_noise(x, noise, timesteps)
show_images((noised_x*0.5 + 0.5).clip(0, 1))

During training, we’ll pick the timesteps at random. The scheduler takes
some parameters (beta_start and beta_end) which it uses to determine how
much noise should be present for a given timestep. We will cover
schedulers in more detail in section X.
The UNet
UNet is a convolutional neural network invented for tasks such as image
segmentation, where the desired output has the same spatial extent as the
input. It consists of a series of ‘downsampling’ layers that reduce the spatial
size of the input, followed by a series of ‘upsampling’ layers that increase
the spatial extent of the input again. The downsampling layers are also
typically followed by a ‘skip connection’ that connects the downsampling
layer’s output to the upsampling layer’s input. This allows the upsampling
layers to ‘see’ the higher-resolution representations from earlier in the
network, which is useful for tasks with image-like outputs where this high-
resolution information is especially useful.
The UNet architecture used in the diffusers library is more advanced than
the original UNet proposed in 2015 by Ronneberger et al, with additions
like attention and residual blocks. We’ll take a closer look later, but the key
feature here is that it can take in an input (the noisy image) and produce a
prediction that is the same shape (the predicted noise). For diffusion
models, the UNet typically also takes in the timestep as additional
conditioning, which again we will explore in the UNet deep dive section.
Here’s how we might create a UNet and feed our batch of noisy images
through it:

# Create a UNet2DModel
model = UNet2DModel(
in_channels=3, # 3 channels for RGB images
sample_size=64, # Specify our input size
block_out_channels=(64, 128, 256, 512), # N channels per
layer
down_block_types=("DownBlock2D", "DownBlock2D",
"AttnDownBlock2D", "AttnDownBlock2D"),
up_block_types=("AttnUpBlock2D", "AttnUpBlock2D",
"UpBlock2D", "UpBlock2D"),
)

# Pass a batch of data through

with torch.no_grad():
out = model(noised_x, timestep=timesteps).sample
out.shape
torch.Size([8, 3, 64, 64])
Note that the output is the same shape as the input, which is exactly what
we want.

Training
Now that we have our model and our data ready, we can train it. We’ll use
the AdamW optimizer with a learning rate of 3e-4. For each training step,
we:
Load a batch of images.
Add noise to the images, choosing random timesteps to determine how
much noise is added.
Feed the noisy images into the model.
Calculate the loss, which is the mean squared error between the
model’s predictions and the target - which in this case is the noise that
we added to the images. This is called the noise or ‘epsilon’ objective.
You can find more information on the different training objectives in
section X.
Backpropagate the loss and update the model weights with the
optimizer.

Here’s what all of that looks like in code:

num_epochs = 50 # How many runs through the data should we do?

lr = 1e-4 # What learning rate should we use
model = model.to(device) # The model we're training (defined in
the previous section)
optimizer = torch.optim.AdamW(model.parameters(), lr=lr) # The
optimizer
losses = [] # somewhere to store the loss values for later
plotting

# Train the model (this takes a while!)

for epoch in range(num_epochs):
for step, batch in enumerate(train_dataloader):

# Load the input images

 clean_images = batch["images"].to(device)

# Sample noise to add to the images

noise =
torch.randn(clean_images.shape).to(clean_images.device)

# Sample a random timestep for each image

timesteps = torch.randint(
0,
scheduler.num_train_timesteps,
(clean_images.shape[0],),
device=clean_images.device,
).long()

# Add noise to the clean images according timestep

noisy_images = scheduler.add_noise(clean_images, noise,
timesteps)

# Get the model prediction for the noise

noise_pred = model(noisy_images, timesteps,
return_dict=False)[0]

# Compare the prediction with the actual noise:

loss = F.mse_loss(noise_pred, noise)

# Store the loss for later plotting

losses.append(loss.item())

# Update the model parameters with the optimizer based on

this loss
loss.backward(loss)
optimizer.step()
optimizer.zero_grad()

It takes an hour or so to run the above code on a GPU, so get some tea
while you wait or lower the number of epochs. Here’s what the loss curve
looks like after training:

# Plot the loss curve:

plt.plot(losses);
The loss curve trends downwards as the model learns to denoise the images.
The curve is fairly noisy, thanks to different amounts of noise being added
to the images based on the random sampling of timesteps for each iteration.
It is hard to tell just by looking at the mean squared error of the noise
predictions whether this model will be any good at generating samples, so
let’s move on to the next section and see how well it does.

Sampling
The diffusers library uses the idea of ‘pipelines’ which bundle together all
of the components needed to generate samples with a diffusion model:

pipeline = DDPMPipeline(unet=model, scheduler=scheduler)

ims = pipeline(batch_size=4).images
show_images(ims, nrows=1)

Of course, offloading the job of creating samples to the pipeline doesn’t

really show us what is going on. So, here is a simple sampling loop that
shows how the model is gradually refining the input image:

# Random starting point (4 random images):

sample = torch.randn(4, 3, 64, 64).to(device)

for i, t in enumerate(scheduler.timesteps):

# Get model pred

with torch.no_grad():
noise_pred = model(sample, t).sample

# Update sample with step

sample = scheduler.step(noise_pred, t, sample).prev_sample

show_images(sample.clip(-1, 1)*0.5 + 0.5, nrows=1)

This is the same code we used at the beginning of the chapter to illustrate
the idea of iterative refinement, but hopefully, now you have a better
understanding of what is going on here. We start with a completely random
input, which is then refined by the model in a series of steps. Each step is a
small update to the input, based on the model’s prediction for the noise at
that timestep. We’re still abstracting away some complexity behind the call
to pipeline.scheduler.step() - in a later chapter we will dive
deeper into different sampling methods and how they work.

Evaluation
Generative model performance can be evaluated using FID scores (Fréchet
Inception Distance). FID scores measure how closely generated samples
match real-world samples by comparing statistics between feature maps
extracted from both sets of data using a pre-trained neural network. The
lower the score, the better the quality and realism of generated images
produced by a given model. FID scores are popular due to their ability to
provide an ‘objective’ comparison metric for different types of generative
networks without relying on human judgment.
As convenient as FID scores are, there are some important caveats to be
aware of:

The FID score for a given model depends on the number of samples
used to calculate it, so when comparing between model,s we need to
make sure both reported scores are calculated using the same number
of samples. Common practice is to use 50,000 samples for this
purpose, although to save time, you may evaluate on a smaller number
of samples during development and only do the full evaluation once
you’re ready to publish the results.
When calculating FID, images are resized to 299px square images.
This makes it less useful as a metric for extremely low-res or high-res
images. There are also minor differences between how resizing is
handled by different deep learning frameworks, which can result in
small differences in the FID score! We recommend using a library such
as clean-fid to standardize the FID calculation.
The network used as a feature extractor for FID is typically a model
trained on the Imagenet classification task. When generating images in
a different domain, the features learned by this model may be less
useful. A more accurate approach would be to somehow train a
classification network on domain-specific data first, but this would
make it harder to compare scores between different papers and
approaches, so for now the imagenet model is the standard choice.
If you save generated samples for later evaluation, the format and
compression can again affect the FID score. Avoid low-quality JPEG
images where possible.
Even if you account for all these caveats, FID scores are just a rough
measure of quality and do not perfectly capture the nuances of what makes
images look more ‘real’. So, use them to get an idea of how one model
performs relative to another but also look at the actual images generated by
each model to get a better sense of how they compare. Human preference is
still the gold standard for quality in what is ultimately a fairly subjective
field!

In Depth: Noise Schedules

In the training example above, one of the steps was ‘add noise, in different
amounts’. We achieved this by picking a random timestep between 0 and
1000 and then relying on the scheduler to add the appropriate amount of
noise. Likewise, during sampling, we again relied on the scheduler to tell us
which timesteps to use and how to move from one to the next given the
model predictions. It turns out that choosing how much noise to add is an
important design decision that can drastically affect the performance of a
given model. In this section, we’ll see why this is the case and explore some
of the different approaches that are used in practice.

Why Add Noise?

At the start of this chapter, we said that the key idea behind diffusion
models is that of iterative refinement. During training, we ‘corrupt’ an input
by different amounts. During inference, we begin with a ‘maximally
corrupted’ input and iteratively ‘de-corrupt’ it, in the hopes that we will
eventually end up with a nice final result.
So far, we’ve focused on one specific kind of ‘corruption’: adding Gaussian
noise. One reason for this is the theoretical underpinnings of diffusion
models - if we use a different corruption method we are no longer
technically doing ‘diffusion’! However, a paper titled Cold Diffusion by
Bansal et al dramatically demonstrated that we do not necessarily need to
constrain ourselves to this method just for theoretical convenience. They
showed that a diffusion-model-like approach works for many different
‘corruption’ methods (see Figure 1-1). More recently, models like MUSE,
MaskGIT and PAELLA have used random token masking or replacement as
an equivalent ‘corruption’ method for quantized data - that is, data that is
represented by discrete tokens rather than continuous values.

Figure 1-1. Illustration of the different degradations used in the Cold Diffusion Paper

Nonetheless, adding noise remains the most popular approach for several
reasons:
We can easily control the amount of noise added, giving a smooth
transition from ‘perfect’ to ‘completely corrupted’. This is not the case
for something like reducing the resolution of an image, which may
result in ‘discrete’ transitions.
We can have many valid random starting points for inference, unlike
some methods which may only have a limited number of possible
initial (fully corrupted) states, such as a completely black image or a
single-pixel image.
So, for the moment at least, we’ll stick with adding noise as our corruption
method. Next, let’s take a closer look at how we add noise to our images.

Starting Simple
We have some images (x) and we’d like to combine them somehow with
some random noise.

x = next(iter(train_dataloader))['images'][:8]
noise = torch.rand_like(x)

One way we could do this is to linearly interpolate (lerp) between them by

some amount. This gives us a function that smoothly transitions from the
original image x to pure noise as the ‘amount’ varies from 0 to 1:

def corrupt(x, noise, amount):

amount = amount.view(-1, 1, 1, 1) # make sure it's
broadcastable
return x*(1-amount) + noise*amount

Let’s see this in action on a batch of data, with the amount of noise varying
from 0 to 1:

amount = torch.linspace(0, 1, 8)
noised_x = corrupt(x, noise, amount)
show_images(noised_x*0.5 + 0.5)
This seems to be doing exactly what we want, smoothly transitioning from
the original image to pure noise. Now, we’ve created a noise schedule here
that takes in a value for ‘amount’ from 0 to 1. This is called the ‘continuous
time’ approach, where we represent the full path on a time scale from 0 to
1. Other approaches use a discrete time approach, with some large integer
number of ‘timesteps’ used to define the noise scheduler. We can wrap our
function into a class that converts from continuous time to discrete
timesteps and adds noise appropriately:

class SimpleScheduler():
def __init__(self):
self.num_train_timesteps = 1000
def add_noise(self, x, noise, timesteps):
amount = timesteps / self.num_train_timesteps
return corrupt(x, noise, amount)

scheduler = SimpleScheduler()
timesteps = torch.linspace(0, 999, 8).long()
noised_x = scheduler.add_noise(x, noise, timesteps)
show_images(noised_x*0.5 + 0.5)
Now we have something that we can directly compare to the schedulers
used in the diffusers library, such as the DDPMScheduler we used during
training. Let’s see how it compares:

scheduler = DDPMScheduler(beta_end=0.01)
timesteps = torch.linspace(0, 999, 8).long()
noised_x = scheduler.add_noise(x, noise, timesteps)
show_images((noised_x*0.5 + 0.5).clip(0, 1))
The Maths
There are many competing notations and approaches in the literature. For
example, some papers parametrize the noise schedule in continuous-time
where t runs from 0 (no noise) to 1 (fully corrupted) - just like our
corrupt function in the previous section. Others use a discrete-time
approach with integer timesteps running from 0 to some large number T,
typically 1000. It is possible to convert between these two approaches the
way we did with our SimpleScheduler class - just make sure you’re
consistent when comparing different models. We’ll stick with the discrete-
time approach here.
A good place to start for pushing deeper into the maths is the DDPM paper
mentioned earlier. You can find an annotated implementation here which is
a great additional resource for understanding this approach.
The paper begins by specifying a single noise step to go from timestep t-1
to timestep t. Here’s how they write it:

q (xt|xt−1) = N (xt; √ 1 − βtxt−1, βtI).

Here β is defined for all timesteps t and is used to specify how much noise
t

is added at each step. This notation can be a little intimidating, but what this
equation tells us is that the noisier x is a distribution with a mean of
t

√1 − β x t and a variance of β . In other words, x is a mix of x

t−1 t t t−1

(scaled by √1 − β ) and some random noise, which we can think of as

t

unit-variance noise scaled by √β . Given x and some noise ϵ, we can

t t−1

sample from this distribution to get x with:t

xt = √ 1 − βtxt−1 + √ βtϵ

To get the noisy input at timestep t, we could begin at t=0 and repeatedly
apply this single step, but this would be very inefficient. Instead, we can
find a formula to move to any timestep t in one go. We define α = 1 − β
t t

and then use the following formula:

xt = √ αtx0 + √ 1 − αtϵ
where - ϵ is some gaussian noise with unit variance - α (‘alpha_bar’) is the
cumulative product of all the α values up to the time t.
So x is a mixture of x (scaled by √α ) and ϵ (scaled by √1 − α ). In the
t 0 t t

diffusers library the α values are stored in

scheduler.alphas_cumprod. Knowing this, we can plot the scaling
factors for the original image x and the noise ϵ across the different
0

timesteps for a given scheduler:

plot_scheduler(DDPMScheduler()) # The default scheduler

Our SimpleScheduler above just linearly mixes between the original image
and noise, as we can see if we plot the scaling factors (equivalent to √αt

and √(1 − α ) in the DDPM case):

t

plot_scheduler(SimpleScheduler())
A good noise schedule will ensure that the model sees a mix of images at
different noise levels. The best choice will differ based on the training data.
Visualizing a few more options, note that:
Setting beta_end too low means we never completely erase the image,
so the model will never see anything like the random noise used as a
starting point for inference.
Setting beta_end extremely high means that most of the timesteps are
spent on almost complete noise, which will result in poor training
performance.
Different beta schedules give different curves.

The ‘cosine’ schedule is a popular choice, as it gives a smooth transition

from the original image to the noise.

fig, (ax) = plt.subplots(1, 1, figsize=(8, 5))

plot_scheduler(DDPMScheduler(beta_schedule="linear"),
label = 'default schedule', ax=ax,
plot_both=False)
plot_scheduler(DDPMScheduler(beta_schedule="squaredcos_cap_v2"),
label = 'cosine schedule', ax=ax, plot_both=False)
plot_scheduler(DDPMScheduler(beta_end=0.003,
beta_schedule="linear"),
label = 'Low beta_end', ax=ax, plot_both=False)
plot_scheduler(DDPMScheduler(beta_end=0.1,
beta_schedule="linear"),
label = 'High beta_end', ax=ax, plot_both=False)
 NOTE
All of the schedules shown here are called ‘Variance Preserving’ (VP), meaning that the
variance of the model input is kept close to 1 across the entire schedule. You may also
encounter ‘Variance Exploding’ (VE) formulations where noise is simply added to the
original image in different amounts (resulting in high-variance inputs). We’ll go into this
more in the chapter on sampling. Our SimpleScheduler is almost a VP schedule, but the
variance is not quite preserved due to the linear interpolation.

As with many diffusion-related topics, there is a constant stream of new

papers exploring the topic of noise schedules, so by the time you read this
there will likely be a large collection of options to try out!

Effect of Input Resolution and Scaling

One aspect of noise schedules that was mostly overlooked until recently is
the effect of input size and scaling. Many papers test potential schedulers on
small-scale datasets and at low resolution, and then use the best-performing
scheduler to train their final models on larger images. The problem with this
is can be seen if we add the same amount of noise to two images of
different sizes.
 Figure 1-2. Comparing the effect of adding noise to images of different sizes

Images at high resolution tend to contain a lot of redundant information.

This means that even if a single pixel is obscured by noise, the surrounding
pixels contain enough information to reconstruct the original image. This is
not the case for low-resolution images, where a single pixel can contain a
lot of information. This means that adding the same amount of noise to a
low-resolution image will result in a much more corrupted image than
adding the equivalent amount of noise to a high-resolution image.
This effect was thoroughly investigated in two independent papers, both of
which came out in January 2023. Each used the new insights to train
models capable of generating high-resolution outputs without requiring any
of the tricks that have previously been necessary. Simple diffusion by
Hoogeboom et al introduced a method for adjusting the noise schedule
based on the input size, allowing a schedule optimized on low-resolution
images to be appropriately modified for a new target resolution. A paper
called “On the Importance of Noise Scheduling for Diffusion Models” by
Ting Chen performed similar experiments, and noted another key variable:
input scaling. That is, how do we represent our images? If the images are
represented as floats between 0 and 1 then they will have a lower variance
than the noise (which is typically unit variance) and thus the signal-to-noise
ratio will be lower for a given noise level than if the images were
represented as floats between -1 and 1 (which we used in the training
example above) or something else. Scaling the input images shifts the
signal-to-noise ratio, and so modifying this scaling is another way we can
adjust when training on larger images.
In Depth: UNets and Alternatives
Now let’s address the actual model that makes the all-important predictions!
To recap, this model must be capable of taking in a noisy image and
estimating how to denoise it. This requires a model that can take in an
image of arbitrary size and output an image of the same size. Furthermore,
the model should be able to make precise predictions at the pixel level,
while also capturing higher-level information about the image as a whole. A
popular approach is to use an architecture called a UNet. UNets were
invented in 2015 for medical image segmentation, and have since become a
popular choice for various image-related tasks. Like the AutoEncoders and
VAEs we looked at in the previous chapter, UNets are made up of a series
of ‘downsampling’ and ‘upsampling’ blocks. The downsampling blocks are
responsible for reducing the size of the image, while the upsampling blocks
are responsible for increasing the size of the image. The downsampling
blocks are typically made up of a series of convolutional layers, followed
by a pooling or downsampling layer. The upsampling blocks are typically
made up of a series of convolutional layers, followed by an upsampling or
‘transposed convolution’ layer. The transposed convolution layer is a
special type of convolutional layer that increases the size of the image,
rather than reducing it.
The reason a regular AutoEncoder or VAE is not a good choice for this task
is that they are less capable of making precise predictions at the pixel level
since the output must be entirely re-constructed from the low-dimensional
latent space. In a UNet, the downsampling and upsampling blocks are
connected by ‘skip connections’, which allow information to flow directly
from the downsampling blocks to the upsampling blocks. This allows the
model to make precise predictions at the pixel level, while also capturing
higher-level information about the image as a whole.

A Simple UNet
To better understand the structure of a UNet, let’s build a simple UNet from
scratch.
 Figure 1-3. Our simple UNet architecture

This UNet takes single-channel inputs at 32px resolution and outputs

single-channel outputs at 32px resolution, which we could use to build a
diffusion model for the MNIST dataset. There are three layers in the
encoding path, and three layers in the decoding path. Each layer consists of
a convolution followed by an activation function and an upsampling or
downsampling step (depending on whether we are in the encoding or
decoding path). The skip connections allow information to flow directly
from the downsampling blocks to the upsampling blocks, and are
implemented by adding the output of the downsampling block to the input
of the corresponding upsampling block. Some UNets instead concatenate
the output of the downsampling block to the input of the corresponding
upsampling block, and may also include additional layers in the skip
connections. Here’s what this network looks like in code:

from torch import nn

class BasicUNet(nn.Module):
"""A minimal UNet implementation."""
def __init__(self, in_channels=1, out_channels=1):
 super().__init__()
self.down_layers = torch.nn.ModuleList([
nn.Conv2d(in_channels, 32, kernel_size=5, padding=2),
nn.Conv2d(32, 64, kernel_size=5, padding=2),
nn.Conv2d(64, 64, kernel_size=5, padding=2),
])
self.up_layers = torch.nn.ModuleList([
nn.Conv2d(64, 64, kernel_size=5, padding=2),
nn.Conv2d(64, 32, kernel_size=5, padding=2),
nn.Conv2d(32, out_channels, kernel_size=5,
padding=2),
])
self.act = nn.SiLU() # The activation function
self.downscale = nn.MaxPool2d(2)
self.upscale = nn.Upsample(scale_factor=2)

def forward(self, x):

h = []
for i, l in enumerate(self.down_layers):
x = self.act(l(x)) # Through the layer and the
activation function
if i < 2: # For all but the third (final) down layer:
h.append(x) # Storing output for skip connection
x = self.downscale(x) # Downscale ready for the
next layer

for i, l in enumerate(self.up_layers):
if i > 0: # For all except the first up layer
x = self.upscale(x) # Upscale
x += h.pop() # Fetching stored output (skip
connection)
x = self.act(l(x)) # Through the layer and the
activation function

return x

A diffusion model trained with this architecture on MNIST produces the

following samples (code included in the supplementary material but omitted
here for brevity):

Improving the UNet

This simple UNet works for this relatively easy task, but it is far from ideal.
So, what can we do to improve it?
 Add more parameters. This can be accomplished by using multiple
convolutional layers in each block, by using a larger number of filters
in each convolutional layer, or by making the network deeper.
Add residual connections. Using ResBlocks instead of regular
convolutional layers can help the model learn more complex functions
while keeping training stable.
Add normalization, such as batch normalization. Batch normalization
can help the model learn more quickly and reliably, by ensuring that
the outputs of each layer are centered around 0 and have a standard
deviation of 1.
Add regularization, such as dropout. Dropout helps by preventing the
model from overfitting to the training data, which is important when
working with smaller datasets.
Add attention. By introducing self-attention layers we allow the model
to focus on different parts of the image at different times, which can
help it learn more complex functions. The addition of transformer-like
attention layers also lets us increase the number of learnable
parameters, which can help the model learn more complex functions.
The downside is that attention layers are much more expensive to
compute than regular convolutional layers at higher resolutions, so we
typically only use them at lower resolutions (i.e. the lower resolution
blocks in the UNet).
Add an additional input for the timestep, so that the model can tailor
its predicitons according to the noise level. This is called timestep
conditioning, and is used in almost all recent diffusion models. We’ll
see more on conditional models in the next chapter.

For comparison, here are the results on MNIST when using the UNet
implementation in the diffusers library, which features all of the above
improvements:
 WARNING
This section will likely be expanded with results and more details in the future. We just
haven’t gotten around to training variants with the different improvements yet!

Alternative Architectures
More recently, a number of alternative architectures have been proposed for
diffusion models. These include:
Transformers. The DiT paper (“Scalable Diffusion Models with
Transformers”) by Peebles and Xie showed that a transformer-based
architecture can be used to train a diffusion model, with great results.
However, the compute and memory requirements of the transformer
architecture remain a challenge for very high resolutions.
The UViT architecture from the Simple Diffusion paper link aims to
get the best of both worlds by replacing the middle layers of the UNet
with a large stack of transformer blocks. A key insight of this paper
was that focusing the majority of the compute at the lower resolution
blocks of the UNet allows for more efficient training of high-resolution
diffusion models. For very high resolutions, they do some additional
pre-processing using something called a wavelet transform to reduce
the spatial resolution of the input image while keeping as much
information as possible through the use of additional channels, again
reducing the amount of compute spent on the higher spatial
resolutions.
Recurrent Interface Networks. The RIN paper (Jabri et al) takes a
similar approach, first mapping the high-resolution inputs to a more
manageable and lower-dimensional ‘latent’ representation which is
then processed by a stack of transformer blocks before being decoded
back out to an image. Additionally, the RIN paper introduces an idea
of ‘recurrence’ where information is passed to the model from the
previous processing step, which can be beneficial for the kind of
iterative improvement that diffusion models are designed to perform.
It remains to be seen whether transformer-based approaches completely
supplant UNets as the go-to architecture for diffusion models, or whether
hybrid approaches like the UViT and RIN architectures will prove to be the
most effective.

In Depth: Objectives and Pre-Conditioning

We’ve spoken about diffusion models taking a noisy input and “learning to
denoise” it. At first glance, you might assume that the natural prediction
target for the network is the denoised version of the image, which we’ll call
x0. However, in the code, we compared the model prediction with the unit-
variance noise that was used to create the noisy version (often called the
epsilon objective, eps). The two appear mathematically identical since if
we know the noise and the timestep we can derive x0 and vice versa. While
this is true, the choice of objective has some subtle effects on how large the
loss is at different timesteps, and thus which noise levels the model learns to
denoise best. To gain some intuition, let’s visualize some different
objectives across different timesteps:

At extremely low noise levels, the x0 objective is trivially easy while

predicting the noise accurately is almost impossible. Likewise, at extremely
high noise levels, the eps objective is easy while predicting the denoised
image accurately is almost impossible. Neither case is ideal, and so
additional objectives have been introduced that have the model predict a
mix of x0 and eps at different timesteps. The v objective (introduced in
“Progressive distillation for fast sampling of diffusion models.” by
Salimans and Ho) is one such objective, which is defined as
v = √ α ⋅ ϵ + √ 1 − α ⋅ x . The EDM paper by Karras et al introduce a
0

similar idea via a parameter called c_skip, and unify the different
diffusion model formulations into a consistent framework. If you’re
interested in learning more about the different objectives, scalings and other
nuances of the different diffusion model formulations, we recommend
reading their paper for a more in-depth discussion.

Project Time: Train Your Own Diffusion

Model
Now that you have an understanding of the basics of diffusion models, it’s
time to train some for yourself! The supplementary material for this chapter
includes a notebook that walks you through the process of training a
diffusion model on your own dataset. As you work through it, check back
with this chapter and see how the different pieces fit together. The notebook
also includes lots of suggested changes you can make to better explore how
different model architectures and training strategies affect the results.

Summary
In this chapter, we’ve seen how the idea of iterative refinement can be
applied to train a diffusion model capable of turning noise into beautiful
images. You’ve seen some of the design choices that go into creating a
successful diffusion model, and hopefully put them into practice by training
your own model. In the next chapter, we’ll take a look at some of the more
advanced techniques that have been developed to improve the performance
of diffusion models and to give them extraordinary new capabilities!

References
Ho, Jonathan, Ajay Jain, and Pieter Abbeel. “Denoising diffusion
probabilistic models.” Advances in Neural Information Processing Systems
33 (2020): 6840-6851.
Ronneberger, O., Fischer, P. and Brox, T., 2015. U-net: Convolutional
networks for biomedical image segmentation. In Medical Image Computing
and Computer-Assisted Intervention–MICCAI 2015: 18th International
Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18
(pp. 234-241). Springer International Publishing.
Bansal, Arpit, Eitan Borgnia, Hong-Min Chu, Jie S. Li, Hamid Kazemi,
Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. “Cold
diffusion: Inverting arbitrary image transforms without noise.” arXiv
preprint arXiv:2208.09392 (2022).
Hoogeboom, Emiel, Jonathan Heek, and Tim Salimans. “simple diffusion:
End-to-end diffusion for high resolution images.” arXiv preprint
arXiv:2301.11093 (2023).
Chang, Huiwen, Han Zhang, Jarred Barber, A. J. Maschinot, Jose Lezama,
Lu Jiang, Ming-Hsuan Yang et al. “Muse: Text-To-Image Generation via
Masked Generative Transformers.” arXiv preprint arXiv:2301.00704
(2023).
Chang, Huiwen, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman.
“Maskgit: Masked generative image transformer.” In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.
11315-11325. 2022.
Rampas, Dominic, Pablo Pernias, Elea Zhong, and Marc Aubreville. “Fast
Text-Conditional Discrete Denoising on Vector-Quantized Latent Spaces.”
arXiv preprint arXiv:2211.07292 (2022).
Chen, Ting “On the Importance of Noise Scheduling for Diffusion Models.”
arXiv preprint arXiv:2301.10972 (2023).
Peebles, William, and Saining Xie. “Scalable Diffusion Models with
Transformers.” arXiv preprint arXiv:2212.09748 (2022).
Jabri, Allan, David Fleet, and Ting Chen. “Scalable Adaptive Computation
for Iterative Generation.” arXiv preprint arXiv:2212.11972 (2022).
Salimans, Tim, and Jonathan Ho. “Progressive distillation for fast sampling
of diffusion models.” arXiv preprint arXiv:2202.00512 (2022).)
Karras, Tero, Miika Aittala, Timo Aila, and Samuli Laine. “Elucidating the
design space of diffusion-based generative models.” arXiv preprint
arXiv:2206.00364 (2022).
Chapter 2. Building up to Stable
Diffusion

A NOTE FOR EARLY RELEASE READERS

With Early Release ebooks, you get books in their earliest form—the
authors’ raw and unedited content as they write—so you can take
advantage of these technologies long before the official release of these
titles.
This will be the fourth chapter of the final book. Please note that the
GitHub repo will be made active later on.
If you have comments about how we might improve the content and/or
examples in this book, or if you notice missing material within this
chapter, please reach out to the editor at [email protected].

In the previous chapter, we introduced diffusion models and the underlying

idea of iterative refinement. By the end of the chapter, we could generate
images, but training the model was time-consuming and we had no control
over the images that were generated. In this chapter, we’ll see how to go
from this to text-conditioned models that can efficiently generate images
based on text descriptions, with a model called Stable Diffusion (SD) as a
case study. Before we get to SD, though, we’ll first look at how conditional
models work and go over some of the innovations that lead up to the text-
to-image models we have today.

Adding Control: Conditional Diffusion

Models
Before we deal with the problem of generating images from text
descriptions (a very challenging task!), let’s focus on something slightly
easier first. We’ll see how we can steer our models outputs towards specific
types or classes of images. We can use a method called conditioning, where
the idea is to ask the model to generate not just any image, but an image
belonging to a pre-defined class.
Model conditioning is a simple but effective idea. We’ll start from the same
diffusion model we used in Chapter 3, with just a couple of changes. First,
we’ll use a new dataset called Fashion MNIST instead of butterflies so that
we can identify categories easily. Then, crucially, we’ll run two inputs
through the model. Instead of just showing it how real images look like,
we’ll also tell it the class every image belongs to. We hope the model will
learn to associate images and labels, so it gets an idea about the distinctive
features of sweaters, boots and the like.
Note that we are not interested in solving a classification problem – we
don’t want the model to tell us the class, given an input image –. We still
want it to perform the same task as in Chapter 3, namely: please, generate
plausible images that look like they came from this dataset. The only
difference is that we are giving it additional information about those
images. We’ll use the same loss function and training strategy, as it’s the
same task as before.

Preparing the Data

We need a dataset with distinct groups of images. Datasets intended for
computer vision classification tasks are ideal for this purpose. We could
start with something like the ImageNet dataset, which contains millions of
images across 1000 classes. However, training models on this dataset would
take an extremely long time. When approaching a new problem, it’s often a
good idea to start with a smaller dataset first, to make sure everything
works as expected. This keeps the feedback loop short, so we can iterate
quickly and make sure we’re on the right track.
For this example, we could choose MNIST as we did in Chapter 3. To make
things just a little bit different, we’ll choose Fashion MNIST instead.
Fashion MNIST, developed and open-sourced by Zalando, is a replacement
for MNIST that shares some of the same characteristics: a compact size,
black & white images, and 10 classes. The main difference is that instead of
being digits, classes correspond to different types of clothing and the
images contain more detail than simple handwritten digits.
Let’s look at some examples.

from datasets import load_dataset

fashion_mnist = load_dataset("fashion_mnist")
clothes = fashion_mnist["train"]["image"][:8]
classes = fashion_mnist["train"]["label"][:8]
show_images(clothes, titles=classes, figsize=(4,2.5))

So class 0 means t-shirt, 2 is a sweater and 9 means boot. Here’s a list of the
10 categories in Fashion MNIST:
https://huggingface.co/datasets/fashion_mnist#data-fields. We prepare our
dataset and dataloader similarly to how we did it in Chapter 3, with the
main difference that we’ll also include the class information as an input.
Instead of resizing, in this case we’ll pad our image inputs (which have a
size of 28 × 28 pixels) to 32 × 32, as we did in Chapter 3.

preprocess = transforms.Compose([
transforms.RandomHorizontalFlip(), # Randomly flip (data
augmentation)
transforms.ToTensor(), # Convert to tensor (0,
1)
transforms.Pad(2), # Add 2 pixels on all
 sides
transforms.Normalize([0.5], [0.5]), # Map to (-1, 1)
])
batch_size = 256

def transform(examples):
images = [preprocess(image.convert("L")) for image in
examples["image"]]
return {"images": images, "labels": examples["label"]}

train_dataset = fashion_mnist["train"].with_transform(transform)

train_dataloader = torch.utils.data.DataLoader(
train_dataset, batch_size=batch_size, shuffle=True
)

Creating a Class-Conditioned Model

If we use the UNet model from the diffusers library, we can provide our
own custom conditioning information. Here we create a similar model to
the one we used in Chapter 3, but we add a num_class_embeds argument to
the UNet constructor. This argument tells the model that we’d like to use
class labels as additional conditioning. We’ll use 10, because we have 10
classes in Fashion MNIST.

model = UNet2DModel(
in_channels=1, # 1 channel for grayscale images
out_channels=1, # output channels must also be 1
sample_size=32,
block_out_channels=(32, 64, 128, 256),
norm_num_groups=8,
num_class_embeds=10, # Enable class conditioning
)

To make predictions with this model, we must pass in the class labels as
additional inputs to the forward method:

x = torch.randn((1, 1, 32, 32))

with torch.no_grad():
out = model(x, timestep=7,
class_labels=torch.tensor([2])).sample
out.shape
torch.Size([1, 1, 32, 32])
 NOTE
You’ll notice we also pass something else to the model as conditioning: the timestep!
That’s right, even the model from Chapter 3 can be considered a conditional diffusion
model! We condition it on the timestep in the hopes that knowing how far we are in the
diffusion process will help it generate more realistic images.

Internally, both the timestep and the class label are turned into embeddings
that the model uses during its forward pass. At multiple stages throughout
the UNet, these embeddings are projected onto a dimension that matches
the number of channels in a given layer and are then added to the outputs of
that layer. This means the conditioning information is fed to every block of
the UNet, giving the model ample opportunity to learn how to use it
effectively.

Training the Model

Adding noise works just as well on greyscale images as it did on the
butterflies from Chapter 3.

scheduler = DDPMScheduler(num_train_timesteps=1000,
beta_start=0.0001, beta_end=0.02)
timesteps = torch.linspace(0, 999, 8).long()
batch = next(iter(train_dataloader))
x = batch['images'][:8]
noise = torch.rand_like(x)
noised_x = scheduler.add_noise(x, noise, timesteps)
show_images((noised_x*0.5 + 0.5).clip(0, 1))
Our training loop is also almost exactly the same as in Chapter 3 too, except
that we now pass the class labels for conditioning. Note that this is just
additional information for the model, but it doesn’t affect our loss function
in any way.

for step, batch in enumerate(train_dataloader):

# Load the input images
clean_images = batch["images"].to(device)
class_labels = batch["labels"].to(device)

# *Sample noise to add to the images*

# *Sample a random timestep for each image*
# *Add noise to the clean images according to the
timestep*

# Get the model prediction for the noise - note the use
of class_labels
noise_pred = model(noisy_images, timesteps,
class_labels=class_labels, return_dict=False)[0]

# *Calculate the loss and update the parameters as

before*
...

In this case we train for 25 epochs - full code can be found in the
supplementary material.
Sampling
Now we have a model that expects two inputs when making predictions: the
image and the class label. We can create samples by beginning with random
noise and then iteratively denoising, passing in whatever class label we’d
like to generate:

def generate_from_class(class_to_generate, n_samples=8):

sample = torch.randn(n_samples, 1, 32, 32).to(device)
class_labels = [class_to_generate] * n_samples
class_labels = torch.tensor(class_labels).to(device)

for i, t in tqdm(enumerate(scheduler.timesteps)):
# Get model pred
with torch.no_grad():
noise_pred = model(sample, t,
class_labels=class_labels).sample

# Update sample with step

sample = scheduler.step(noise_pred, t,
sample).prev_sample

return sample.clip(-1, 1)*0.5 + 0.5

# Generate t-shirts (class 0)
images = generate_from_class(0)
show_images(images, nrows=2)
1000it [00:21, 47.25it/s]
# Now generate some sneakers (class 7)
images = generate_from_class(7)
show_images(images, nrows=2)
1000it [00:21, 47.20it/s]

# ...or boots (class 9)

images = generate_from_class(9)
show_images(images, nrows=2)
1000it [00:21, 47.26it/s]
As you can see, the generated images are far from perfect. They’d probably
get much better if we explored the architecture and trained for longer. But
it’s amazing that the model not only learnt the shapes of different types of
clothing, but also realized that shape 9 looks different than shape 0, just by
sending this information alongside the training data. To put it in a slightly
different way: the model is used to seeing the number 9 accompanying
boots. When we ask it to generate an image and provide the 9, it responds
with a boot. We have successfully built a class-conditioned model capable
of generating images conditioned on class labels from fasionMNIST!

Improving Efficiency: Latent Diffusion

Now that we can train a conditional model, all we need to do is scale it up
and condition it on text instead of class labels, right? Well, not quite. As
image size grows, so does the computational power required to work with
those images. This is especially pronounced in an operation called self-
attention, where the amount of operations grows quadratically with the
number of inputs. A 128px square image has 4x as many pixels as a 64px
square image, and so requires 16x (i.e. 4 ) the memory and compute in a
2

self-attention layer. This is a problem for anyone who’d like to generate

high-resolution images!

Figure 2-1. The architecture introduced in the Latent Diffusion Models paper. Note the VAE encoder
and decoder on the left for translating between pixel space and latent space
Latent diffusion tries to mitigate this issue by using a separate model called
a Variational Auto-Encoder (VAE). As we saw in Chapter 2, VAEs can
compress images to a smaller spatial dimension. The rationale behind this is
that images tend to contain a large amount of redundant information - given
enough training data, a VAE can hopefully learn to produce a much smaller
representation of an input image and then reconstruct the image based on
this small latent representation with a high degree of fidelity. The VAE used
in SD takes in 3-channel images and produces a 4-channel latent
representation with a reduction factor of 8 for each spatial dimension. That
is, a 512px square input image will be compressed down to a 4x64x64
latent.
By applying the diffusion process on these smaller latent representations
rather than on full-resolution images, we can get many of the benefits that
would come from using smaller images (lower memory usage, fewer layers
needed in the UNet, faster generation times…) and still decode the result
back to a high-resolution image once we’re ready to view it. This
innovation dramatically lowers the cost to train and run these models. The
paper that introduced this idea (High-Resolution Image Synthesis with
Latent Diffusion Models by Rombach et al) demonstrated the power of this
technique by training models conditioned on segmentation maps, class
labels and text. The impressive results led to further collaboration between
the authors and partners such as RunwayML, LAION, and EleutherAI to
train a more powerful version of the model, which later became Stable
Diffusion.

Stable Diffusion: Components in Depth

Stable Diffusion is a text-conditioned latent diffusion model. Thanks to its
popularity, there are hundreds of websites and apps that let you use it to
create images with no technical knowledge required. It’s also very well-
supported by libraries like diffusers, which let us sample an image with
SD using a user-friendly pipeline:

pipe("Watercolor illustration of a rose").images[0]

 0%| | 0/50 [00:00<?, ?it/s]

In this section we’ll explore all of the components that make this possible.

The Text Encoder

So how does Stable Diffusion understand text? Earlier on we showed how
feeding additional information to the UNet allows us to have some
additional control over the types of images generated. Given a noisy version
of an image, the model is tasked with predicting the denoised version based
on additional clues such as a class label. In the case of SD, the additional
clue is the text prompt. At inference time, we can feed in the description of
an image we’d like to see and some pure noise as a starting point, and the
model does its best to denoise the random input into something that
matches the caption.

Figure 2-2. The text encoder turns an input string into text embeddings which are fed into the UNet
along with the timestep and the noisy latents.

For this to work, we need to create a numeric representation of the text that
captures relevant information about what it describes. To do this, SD
leverages a pre-trained transformer model based on CLIP, which was also
introduced in Chapter 2. The text encoder is a transformer model that takes
in a sequence of tokens and produces a 1024-dimensional vector for each
token (0r 768-dimensional in the case of SD version 1 which we’re using
for the demonstrations in this section). Instead of combining these vectors
into a single representation, we keep them separate and use them as
conditioning for the UNet. This allows the UNet to make use of the
information in each token separately, rather than just the overall meaning of
the entire prompt. Because we’re extracting these text embeddings from the
internal representation of the CLIP model, they are often called the
“encoder hidden states”. Figure 3 shows the text encoder architecture.
Figure 2-3. Diagram showing the text encoding process which transforms the input prompt into a set
of text embeddings (the encoder_hidden_states) which can then be fed in as conditioning to the UNet.

The first step to encode text is to follow a process called tokenization. This
converts a sequence of characters into a sequence of numbers, where each
number represents a group of various characters. Characters that are usually
found together (like most common words) can be assigned a single token
that represents the whole word or group. Long or complicated words, or
words with many inflections, may be translated to multiple tokens, where
each one usually represents a meaningful section of the word.
There is no single “best” tokenizer; instead, each language model comes
with its own one. Differences reside in the number of tokens supported, and
on the tokenization strategy – do we use single characters, as we just
described, or should we consider different primitive units. In the following
example we see how the tokenization of a phrase works with Stable
Diffusion’s tokenizer. Each word in our sentence is assigned a unique token
number (for example, photograph happens to be 8853 in the tokenizer’s
vocabulary). There are also additional tokens that are used to provide
additional context, such as the point where the sentence ends.

prompt = 'A photograph of a puppy'

# Turn the text into a sequence of tokens:
text_input = pipe.tokenizer(prompt, padding="max_length",

max_length=pipe.tokenizer.model_max_length,
truncation=True, return_tensors="pt")

# See the individual tokens

for t in text_input['input_ids'][0][:8]: # We'll just look at the
first 7
print(t, pipe.tokenizer.decoder.get(int(t)))
tensor(49406) <|startoftext|>
tensor(320) a</w>
tensor(8853) photograph</w>
tensor(539) of</w>
tensor(320) a</w>
tensor(6829) puppy</w>
tensor(49407) <|endoftext|>
tensor(49407) <|endoftext|>

Once the text is tokenized, we can pass it through the text encoder to get the
final text embeddings that will be fed into the UNet:

# Grab the output embeddings

text_embeddings =
pipe.text_encoder(text_input.input_ids.to(device))[0]
print('Text embeddings shape:', text_embeddings.shape)
Text embeddings shape: torch.Size([1, 77, 768])

We’ll go into more detail about how a transformer model processes a string
of tokens in the chapters focusing on transformer models.
Classifier-free guidance
It turns out that even with all of the effort put into making the text
conditioning as useful as possible, the model still tends to default to relying
mostly on the noisy input image rather than the prompt when making its
predictions. In a way, this makes sense - many captions are only loosely
related to their associated images and so the model learns not to rely too
heavily on the descriptions! However, this is undesirable when it comes
time to generate new images - if the model doesn’t follow the prompt then
we may get images out that don’t relate to our description at all.

Figure 2-4. Images generated from the prompt “An oil painting of a collie in a top hat” with CFG
scale 0, 1, 2 and 10 (left to right)

To fix this, we use a trick called Classifier-Free Guidance (CGF). During

training, text conditioning is sometimes kept blank, forcing the model to
learn to denoise images with no text information whatsoever (unconditional
generation). Then at inference time, we make two separate predictions: one
with the text prompt as conditioning and one without. We can then use the
difference between these two predictions to create a final combined
prediction that pushes even further in the direction indicated by the text-
conditioned prediction according to some scaling factor (the guidance
scale), hopefully resulting in an image that better matches the prompt.
Figure 4 shows the outputs for a prompt at different guidance scales - as
you can see, higher values result in images that better match the description.

The VAE
The VAE is tasked with compressing images into a smaller latent
representation and back again. The VAE used with Stable Diffusion is a
truly impressive model. We won’t go into the training details here, but in
addition to the usual reconstruction loss and KL divergence described in
Chapter 2 they use an additional patch-based discriminator loss to help the
model learn to output plausible details and textures. This adds a GAN-like
component to training and helps to avoid the slightly blurry outputs that
were typical in previous VAEs. Like the text encoder, the VAE is usually
trained separately and used as a frozen component during the diffusion
model training and sampling process.

Figure 2-5. Encoding and decoding and image with the VAE

Let’s load an image and see what it looks like after being compressed and
decompressed by the VAE:

# NB, this will be our own image as part of the supplementary

material to avoid external URLs
im =
load_image('https://images.pexels.com/photos/14588602/pexels-
photo-14588602.jpeg', size=(512, 512))
show_image(im);
 # Encode the image
with torch.no_grad():
tensor_im = transforms.ToTensor()
(im).unsqueeze(0).to(device)*2-1
latent = vae.encode(tensor_im.half()) # Encode the image to a
distribution
latents = latent.latent_dist.sample() # Sampling from the
distribution
latents = latents * 0.18215 # This scaling factor was
introduced by the SD authors to reduce the variance of the
latents

latents.shape
torch.Size([1, 4, 64, 64])

Visualizing the low-resolution latent representation, we can that some of the

rough structure of the input image is still visible in the different channels:

# Plot the individual channels of the latent representation

show_images([l for l in latents[0]], titles=[f'Channel {i}' for i
in range(latents.shape[1])], ncols=4)

And decoding back to image space, we get an output image that is almost
identical to the original. Can you spot the difference?

# Decode the image

with torch.no_grad():
image = vae.decode(latents / 0.18215).sample
image = (image / 2 + 0.5).clamp(0, 1)
show_image(image[0].float());
When generating images from scratch, we create a random set of latents as
the starting point. We iteratively refine these noisy latents to generate a
sample, and then the VAE decoder is used to decode these final latents into
an image we can view. The encoder is only used if we’d like to start the
process from an existing image, something we’ll explore in chapter 5.

The UNet
The UNet used in stable diffusion is somewhat similar to the one we used in
chapter 3 for generating images. Instead of taking in a 3-channel image as
the input we take in a 4-channel latent. The timestep embedding is fed in in
the same way as the class conditioning was in the example at the start of
this chapter. But this UNet also needs to accept the text embeddings as
additional conditioning. Scattered throughout the UNet are cross-attention
layers. Each spatial location in the UNet can attend to different tokens in
the text conditioning, bringing in relevant information from the prompt. The
diagram in Figure 7 shows how this text conditioning (as well as timestep-
based conditioning) is fed in at different points.
 Figure 2-6. The Stable Diffusion UNet

The UNet for Stable Diffusion version 1 and 2 has around 860 million
parameters. The more recent SD XL has even more, at around (details
TBC), with most of the additional parameters being added at the lower-
resolution stages via additional channels in the residual blocks (N vs 1280
in the original) and additional transformer blocks.
NB: Stable Diffusion XL has not yet been publically released, so this
section will be updated when more information is public.

Putting it All Together: Annotated Sampling

Loop
Now that we know what each of the components does, let’s put them
together to generate an image without relying on the pipeline. Here are the
settings we’ll use:

# Some settings
prompt = ["Acrylic palette knife painting of a flower"] # What we
want to generate
height = 512 # default height of Stable
Diffusion
width = 512 # default width of Stable
 Diffusion
num_inference_steps = 30 # Number of denoising steps
guidance_scale = 7.5 # Scale for classifier-free
guidance
seed = 42 # Seed for random number
generator

The first step is to encode the text prompt. Because we plan to do classifier-
free guidance, we’ll actually create two sets of text embeddings: one with
the prompt, and one representing an empty string. You can also encode a
negative prompt in place of the empty string, or combine multiple prompts
with different weightings, but this is the most common usage:

# Tokenize the input

text_input = pipe.tokenizer(prompt, padding="max_length",
max_length=pipe.tokenizer.model_max_length, truncation=True,
return_tensors="pt")

# Feed through the text encoder

with torch.no_grad():
text_embeddings =
pipe.text_encoder(text_input.input_ids.to(device))[0]

# Do the same for the unconditional input (a blank string)

uncond_input = pipe.tokenizer("", padding="max_length",
max_length=pipe.tokenizer.model_max_length, return_tensors="pt")
with torch.no_grad():
uncond_embeddings =
pipe.text_encoder(uncond_input.input_ids.to(device))[0]

# Concatenate the two sets of text embeddings embeddings

text_embeddings = torch.cat([uncond_embeddings, text_embeddings])

Next we create our random initial latents and set up the scheduler to use the
desired number of inference steps:

# Prepare the Scheduler

pipe.scheduler.set_timesteps(num_inference_steps)

# Prepare the random starting latents

latents = torch.randn(
(1, pipe.unet.in_channels, height // 8, width // 8), # Shape
of the latent representation
 generator=torch.manual_seed(32), # Seed the random number
generator
).to(device).half()
latents = latents * pipe.scheduler.init_noise_sigma

Now we loop through the sampling steps, getting the model prediction at
each stage and using this to update the latents:

# Sampling loop
for i, t in enumerate(pipe.scheduler.timesteps):

# Create two copies of the latents to match the two text

embeddings (unconditional and conditional)
latent_model_input = torch.cat([latents] * 2)
latent_model_input =
pipe.scheduler.scale_model_input(latent_model_input, t)

# predict the noise residual for both sets of inputs

with torch.no_grad():
noise_pred = pipe.unet(latent_model_input, t,
encoder_hidden_states=text_embeddings).sample

# Split the prediction into unconditional and conditional

versions:
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)

# perform classifier-free guidance

noise_pred = noise_pred_uncond + guidance_scale *
(noise_pred_text - noise_pred_uncond)

# compute the previous noisy sample x_t -> x_t-1

latents = pipe.scheduler.step(noise_pred, t,
latents).prev_sample

Notice the classifier-free guidance step. Our final noise prediction is

noise_pred_uncond + guidance_scale * (noise_pred_text -
noise_pred_uncond), pushing the prediction away from the unconditional
prediction towards the prediction made based on the prompt. Try changing
the guidance scale to see how this affects the output.
By the end of the loop the latents should hopefully now represent a
plausible image that matches the prompt. The final step is to decode the
latents into an image using the VAE so that we can see the result:
 # scale and decode the image latents with vae
latents = 1 / 0.18215 * latents
with torch.no_grad():
image = vae.decode(latents).sample
image = (image / 2 + 0.5).clamp(0, 1)

# Display
show_image(image[0].float());

If you explore the source code for the StableDiffusionPipeline you’ll see
that the code above closely matches the call method used by the pipeline.
Hopefully this annotated version shows that there is nothing too magical
going on behind the scenes! Use this as a reference for when we encounter
additional pipelines that add additional tricks to this foundation.

Training Data for Text-To-Image Models

(TBD)
NB: We may add a more in-depth section here with the history and
technical details of how LAION came together, and some of the nuances
and debate around training on public data scraped from the internet.

Open Data, Open Models

The LAION-5B dataset includes over 5 billion image-caption pairs scraped
from the internet. This dataset was created by and for the open-source
community, which saw the need for a publically-accessible dataset of this
kind. Before the LAION initiative, only a handful of research labs at large
companies had access to such data. These organizations kept the details of
their private datasets to themselves, which made their results impossible to
validate or replicate. By creating a publically available source of training
data, LAION enabled a wave of smaller communities and organizations to
train models and perform research that would otherwise have been
impossible.

Figure 2-7. “An explosion of artistic creativity” - Image generated by the authors using Stable
Diffusion

Stable Diffusion was one such model, trained on a subset of LAION as part
of a collaboration between the researchers who had invented latent diffusion
models and an organization called Stability AI. Training a model like SD
requires a significant amount of GPU time. Even with the freely-available
LAION dataset, there aren’t many who could afford the investment. This is
why the public release of the model weights and code was such a big deal -
it marked the first time a powerful text-to-image model with similar
capabilities to the best closed-source alternatives was available to all. Stable
Diffusion’s public availability has made it the go-to choice for researchers
and developers looking to explore this technology over the past year.
Hundreds of papers build upon the base model, adding new capabilities or
finding innovative ways to improve its speed and quality. And innumerable
startups have found ways to integrate these rapidly-improving tools into
their products, spawning an entire ecosystem of new applications.
The months after the introduction of Stable Diffusion demonstrated the
impact of sharing these technologies in the open. SD is not the best text-to-
image model, but it IS the best model most of us had access to, so
thousands of people have spent their time making it better and building
upon that open foundation. We hope this example encourages others to
follow suit and share their work with the open-source community in the
future!

Summary
In this chapter we’ve seen how conditioning gives us new ways to control
the images generated by diffusion models. We’ve seen how latent diffusion
lets us train diffusion models more efficiently. We’ve seen how a text
encoder can be used to condition a diffusion model on a text prompt,
enabling powerful text-to-image capabilities. And we’ve explored how all
of this comes together in the Stable Diffusion model by digging into the
sampling loop and seeing how the different components work together. In
the next chapter, we’ll show some of the many additional capabilities that
can be added to diffusion models such as SD to take them beyond simple
image generation. And later, in part 2 of the book, you’ll learn how to fine-
tune SD to add new knowledge or capabilities to the model.
About the Authors
Pedro Cuenca is a machine learning engineer who works on diffusion
software, models, and applications at Hugging Face.
Apolinário Passos is a machine learning art engineer at Hugging Face,
working across teams on multiple machine learning for art and creativity
use cases.
Omar Sanseviero is a lead machine learning engineer at Hugging Face,
where he works at the intersection of open source, community, and product.
Previously, Omar worked at Google on Google Assistant and TensorFlow.
Jonathan Whitaker is a data scientist and deep learning researcher focused
on generative modeling. Besides his research and consulting work, his main
focus is on sharing knowledge, which he does via the DataScienceCastnet
YouTube channel and various free online resources he has created.