Cascaded Diffusion Models

for High Fidelity Image Generation

Jonathan Ho∗ [email protected]

Chitwan Saharia∗ [email protected]
William Chan [email protected]
David J. Fleet
arXiv:2106.15282v3 [cs.CV] 17 Dec 2021

[email protected]
Mohammad Norouzi [email protected]
Tim Salimans [email protected]
Google, 1600 Amphitheatre Parkway, Mountain View, CA 94043

Abstract
We show that cascaded diffusion models are capable of generating high fidelity images on
the class-conditional ImageNet generation benchmark, without any assistance from auxiliary
image classifiers to boost sample quality. A cascaded diffusion model comprises a pipeline
of multiple diffusion models that generate images of increasing resolution, beginning with a
standard diffusion model at the lowest resolution, followed by one or more super-resolution
diffusion models that successively upsample the image and add higher resolution details.
We find that the sample quality of a cascading pipeline relies crucially on conditioning
augmentation, our proposed method of data augmentation of the lower resolution condi-
tioning inputs to the super-resolution models. Our experiments show that conditioning
augmentation prevents compounding error during sampling in a cascaded model, helping
us to train cascading pipelines achieving FID scores of 1.48 at 64×64, 3.52 at 128×128
and 4.88 at 256×256 resolutions, outperforming BigGAN-deep, and classification accuracy
scores of 63.02% (top-1) and 84.06% (top-5) at 256×256, outperforming VQ-VAE-2.
Keywords: generative models, diffusion models, score matching, iterative refinement,
super-resolution
256×256

64×64
32×32
Class ID = 213
“Irish Setter”

Model 1 Model 2 Model 3

Figure 1: A cascaded diffusion model comprising a base model and two super-resolution models.

∗. Equal contribution

©2021 Jonathan Ho, Chitwan Saharia, William Chan, David J. Fleet, Mohammad Norouzi, Tim Salimans.
License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/.
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

Figure 2: Selected synthetic 256×256 ImageNet samples.

1. Introduction
Diffusion models (Sohl-Dickstein et al., 2015) have recently been shown to be capable of
synthesizing high quality images and audio (Chen et al., 2021; Ho et al., 2020; Kong et al.,
2021; Song and Ermon, 2020): an application of machine learning that has long been
dominated by other classes of generative models such as autoregressive models, GANs, VAEs,
and flows (Brock et al., 2019; Dinh et al., 2017; Goodfellow et al., 2014; Ho et al., 2019;
Kingma and Dhariwal, 2018; Kingma and Welling, 2014; Razavi et al., 2019; van den Oord
et al., 2016a,b, 2017). Most previous work on diffusion models demonstrating high quality
samples has focused on data sets of modest size, or data with strong conditioning signals.
Our goal is to improve the sample quality of diffusion models on large high-fidelity data sets
for which no strong conditioning information is available. To showcase the capabilities of
the original diffusion formalism, we focus on simple, straightforward techniques to improve
the sample quality of diffusion models; for example, we avoid using extra image classifiers to
boost sample quality metrics (Dhariwal and Nichol, 2021; Razavi et al., 2019).
Our key contribution is the use of cascades to improve the sample quality of diffusion
models on class-conditional ImageNet . Here, cascading refers to a simple technique to model
high resolution data by learning a pipeline of separately trained models at multiple resolutions;
a base model generates low resolution samples, followed by super-resolution models that
upsample low resolution samples into high resolution samples. Sampling from a cascading
pipeline occurs sequentially, first sampling from the low resolution base model, followed by
sampling from super-resolution models in order of increasing resolution. While any type of
generative model could be used in a cascading pipeline (e.g., Menick and Kalchbrenner, 2019;
Razavi et al., 2019), here we restrict ourselves to diffusion models. Cascading has been shown
in recent prior work to improve the sample quality of diffusion models (Saharia et al., 2021;
Nichol and Dhariwal, 2021); our work here concerns the improvement of diffusion cascading
pipelines to attain the best possible sample quality.

2
 Cascaded Diffusion Models

The simplest and most effective technique we found to improve cascading diffusion
pipelines is to apply strong data augmentation to the conditioning input of each super-
resolution model. We refer to this technique as conditioning augmentation. In our experiments,
conditioning augmentation is crucial for our cascading pipelines to generate high quality
samples at the highest resolution. With this approach we attain FID scores on class-
conditional ImageNet generation that are better than BigGAN-Deep (Brock et al., 2019) at any
truncation value, and classification accuracy scores that are better than VQ-VAE-2 (Razavi
et al., 2019). We empirically find that conditioning augmentation is effective because it
alleviates compounding error in cascading pipelines due to train-test mismatch, sometimes
referred to as exposure bias in the sequence modeling literature (Bengio et al., 2015; Ranzato
et al., 2016).
The key contributions of this paper are as follows:

• We show that our Cascaded Diffusion Models (CDM) yield high fidelity samples
superior to BigGAN-deep (Brock et al., 2019) and VQ-VAE-2 (Razavi et al., 2019)
in terms of FID score (Heusel et al., 2017) and classification accuracy score (Ravuri
and Vinyals, 2019), the latter by a large margin. We achieve these results with pure
generative models that are not combined with any classifier.

• We introduce conditioning augmentation for our super-resolution models, and find it

critical towards achieving high sample fidelity. We perform an in-depth exploration
of augmentation policies, and find Gaussian augmentation to be a key ingredient for
low resolution upsampling, and Gaussian blurring for high resolution upsampling. We
also show how to efficiently train models amortized over varying levels of conditioning
augmentation to enable post-training hyperparameter search for optimal sample quality.

Section 2 reviews recent work on diffusion models. Section 3 describes the most effective
types of conditioning augmentation that we found for class-conditional ImageNet generation.
Section 4 contains our sample quality results, ablations, and experiments on additional
datasets. Appendix A contains extra samples and Appendix B contains details on hyperpa-
rameters and architectures. High resolution figures and additional supplementary material
can be found at https://cascaded-diffusion.github.io/.

2. Background
We begin with background on diffusion models, their extension to conditional generation,
and their associated neural network architectures.

2.1 Diffusion Models

A diffusion model (Sohl-Dickstein et al., 2015; Ho et al., 2020) is defined by a forward process
that gradually destroys data x0 ∼ q(x0 ) over the course of T timesteps

T
Y p
q(x1:T |x0 ) = q(xt |xt−1 ), q(xt |xt−1 ) = N (xt ; 1 − βt xt−1 , βt I)
t=1

3
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

and a parameterized reverse process pθ (x0 ) = pθ (x0:T ) dx1:T , where

R

T
Y
pθ (x0:T ) = p(xT ) pθ (xt−1 |xt ), pθ (xt−1 |xt ) = N (xt−1 ; µθ (xt , t), Σθ (xt , t)).
t=1

The forward process hyperparameters βt are set so that xT is approximately distributed

according to a standard normal distribution, so p(xT ) is set to a standard normal prior as
well. The reverse process is trained to match the joint distribution of the forward process by
optimizing the evidence lower bound (ELBO) −Lθ (x0 ) ≤ log pθ (x0 ):
" #
X
Lθ (x0 ) = Eq LT (x0 ) + DKL (q(xt−1 |xt , x0 ) k pθ (xt−1 |xt )) − log pθ (x0 |x1 ) (1)
t>1

where LT (x0 ) = DKL (q(xT |x0 ) k p(xT )). The forward process posteriors q(xt−1 |xt , x0 ) and
marginals q(xt |x0 ) are Gaussian, and the KL divergences in the ELBO can be calculated in
closed form. Thus it is possible to train the diffusion model by taking stochastic gradient
steps on random terms of Eq. (1). As previously suggested (Ho et al., 2020; Nichol and
Dhariwal, 2021), we use the reverse process parameterizations
 
1 βt
µθ (xt , t) = √ xt − √ θ (xt , t)
αt 1 − ᾱt
Σii i
θ (xt , t) = exp(log β̃t + (log βt − log β̃t )vθ (xt , t))

where αt = 1 − βt , ᾱt = ts=1 αs , and β̃t = 1− ᾱt−1

1−ᾱt βt .
Q
Sample quality can be improved, at the cost of log likelihood, by optimizing modified
losses instead of the ELBO. The particular form of the modified loss depends on whether we
are learning Σθ or treating it as a fixed hyperparameter (and whether Σθ is learned is itself
considered a hyperparameter choice that we set experimentally). For the case of non-learned
Σθ , we use the simplified loss
h √ √ 2
i
Lsimple (θ) = Ex0 ,∼N (0,I),t∼U ({1,...,T }) θ ( ᾱt x0 + 1 − ᾱt , t) − 

which is a weighted form of the ELBO that resembles denoising score matching over multiple
noise scales (Ho et al., 2020; Song and Ermon, 2019). For the case of learned Σθ , we employ
a hybrid loss (Nichol and Dhariwal, 2021) implemented using the expression

Lhybrid (θ) = Lsimple (θ) + λLvb (θ)

where Lvb = Ex0 [Lθ (x0 )] and a stop-gradient is applied to the θ term inside Lθ . Optimizing
this hybrid loss has the effect of simultaneously learning µθ using Lsimple and learning Σθ
using the ELBO.

2.2 Conditional Diffusion Models

In the conditional generation setting, the data x0 has an associated conditioning signal c, for
example a label in the case of class-conditional generation, or a low resolution image in the
case of super-resolution (Saharia et al., 2021; Nichol and Dhariwal, 2021). The goal is then

4
 Cascaded Diffusion Models

to learn a conditional model pθ (x0 |c). To do so, we modify the diffusion model to include c
as input to the reverse process:
T
Y
pθ (x0:T |c) = p(xT ) pθ (xt−1 |xt , c), pθ (xt−1 |xt , c) = N (xt−1 ; µθ (xt , t, c), Σθ (xt , t, c))
t=1
" #
X
Lθ (x0 |c) = Eq LT (x0 ) + DKL (q(xt−1 |xt , x0 ) k pθ (xt−1 |xt , c)) − log pθ (x0 |x1 , c) .
t>1

The data and conditioning signal (x0 , c) are sampled jointly from the data distribution,
now called q(x0 , c), and the forward process q(x1:T |x0 ) remains unchanged. The only
modification that needs to be made is to inject c as a extra input to the neural network
function approximators: instead of µθ (xt , t) we now have µθ (xt , t, c), and likewise for Σθ .
The particular architectural choices for injecting these extra inputs depends on the type of
the conditioning c, as described next.

2.3 Architectures
The current best architectures for image diffusion models are U-Nets (Ronneberger et al., 2015;
Salimans et al., 2017), which are a natural choice to map corrupted data xt to reverse process
parameters (µθ , Σθ ) that have the same spatial dimensions as xt . Scalar conditioning, such
as a class label or a diffusion timestep t, is provided by adding embeddings into intermediate
layers of the network (Ho et al., 2020). Lower resolution image conditioning is provided
by channelwise concatenation of the low resolution image, processed by bilinear or bicubic
upsampling to the desired resolution, with the reverse process input xt , as in the SR3 (Saharia
et al., 2021) and Improved DDPM (Nichol and Dhariwal, 2021) models. See Fig. 3 for an
illustration of the SR3-based architecture that we use in this work.

3. Conditioning Augmentation in Cascaded Diffusion Models

Suppose x0 is high resolution data and z0 is its low resolution counterpart. We use the
term cascading pipeline to refer to a sequence of generative models. At the low resolution
we have a diffusion model pθ (z0 ), and at the high resolution, a super-resolution diffusion
model pθ (x0 |z0 ). TheR cascading pipeline forms a latent variable model for high resolution
data; i.e., pθ (x0 ) = pθ (x0 |z0 )pθ (z0 ) dz0 . It is straightforward to extend this to more than
two resolutions. It is also straightforward to condition an entire cascading pipeline on class
information or other conditioning information c: the models take on the form pθ (z0 |c) and
pθ (x0 |z0 , c), each using the conditioning mechanism described in Section 2.2. An example
cascading pipeline is depicted in Fig. 4.
Cascading pipelines have been shown to be useful with other generative model fami-
lies (Menick and Kalchbrenner, 2019; Razavi et al., 2019). A major benefit to training a
cascading pipeline over training a standard model at the highest resolution is that most of
the modeling capacity can be dedicated to low resolutions, which empirically are the most
important for sample quality, and training and sampling at low resolutions tends to be the
most computationally efficient. In addition, cascading allows the individual models to be
trained independently, and architecture choices can be tuned at each specific resolution for
the best performance of the entire pipeline.

5
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

) , MK
N 2
(K ) , MK
N 2
(K (K ) , 2 × MK
N 2

( N2 )2 , M2 ( N2 )2 , 2 × M2

(xt , z) N 2 , M1 N 2 , 2 × M1 xt−1

Figure 3: The U-Net architecture used in each model of a CDM pipeline. The first model is a
class-conditional diffusion model that receives the noisy image xt and the class label
y and as input. (The class label y and timestep t are injected into each block as an
embedding, not depicted here). The remaining models in the pipeline are class-conditional
super-resolution models that receive xt , y, and an additional upsampled low-resolution
image z as input. The downsampling/upsampling blocks adjust the image input resolution
N × N by a factor of 2 through each of the K blocks. The channel count at each block
is specified using channel multipliers M1 , M2 , ..., MK , and the upsampling pass has
concatenation skip connections to the downsampling pass.
256×256

64×64
32×32
Class ID = 933
“Cheeseburger” Class Class
Class
Conditional Conditional
Conditional
Super-Res Super-Res

Figure 4: Detailed CDM pipeline for generation of class conditional 256×256 images. The first
model is a class-conditional diffusion model, and it is followed by a sequence of two
class-conditional super-resolution diffusion models. Each model has a U-Net architecture
as depicted in Fig. 3.

The most effective technique we found to improve the sample quality of cascading pipelines
is to train each super-resolution model using data augmentation on its low resolution input.
We refer to this general technique as conditioning augmentation. At a high level, for some
super-resolution model pθ (x0 |z) from a low resolution image z to a high resolution image x0 ,
conditioning augmentation refers to applying some form of data augmentation to z. This
augmentation can take any form, but what we found most effective at low resolutions is
adding Gaussian noise (forward process noise), and for high resolutions, randomly applying
Gaussian blur to z. In some cases, we found it more practical to train super-resolution
models amortized over the strength of conditioning augmentation and pick the best strength
in a post-training hyperparameter search for optimal sample quality. Details on conditioning

6
 Cascaded Diffusion Models

augmentation and its realization during training and sampling are given in the following
sections.

3.1 Blurring Augmentation

One simple instantiation of conditioning augmentation is augmentation of z by blurring.
We found this to be most effective for upsampling to images with resolution 128×128 and
256×256. More specifically, we apply a Gaussian filter of size k and sigma σ to obtain zb .
We use a filter size of k = (3, 3) and randomly sample σ from a fixed range during training.
We perform hyper-parameter search to find the range for σ. During training, we apply this
blurring augmentation to 50% of the examples. During inference, no augmentation is applied
to low resolution inputs. We explored applying different amounts of blurring augmentations
during inference, but did not find it helpful in initial experiments.

3.2 Truncated Conditioning Augmentation

Here we describe what we call truncated conditioning augmentation, a form of conditioning
augmentation that requires a simple modification to the training and architecture of the
super-resolution models, but no change to the low resolution model at the initial stage of
the cascade. We found this method to be most useful at resolutions smaller than 128×128.
Normally, generating a high resolution sample x0 involves first generating z0 from the low
resolution model pθ (z0 ), then feeding that result into the super-resolution model pθ (x0 |z0 ).
In other words, generating a high resolution sample is performed using ancestral sampling
from the latent variable model
Z Z
pθ (x0 ) = pθ (x0 |z0 )pθ (z0 ) dz0 = pθ (x0 |z0 )pθ (z0:T ) dz0:T .

(For simplicity, we have assumed that the low resolution and super-resolution models both use
the same number of timesteps T .) Truncated conditioning augmentation refers to truncating
the low resolution reverse process to stop at timestep s > 0, instead of 0; i.e.,
Z Z
psθ (x0 ) = pθ (x0 |zs )pθ (zs ) dzs = pθ (x0 |zs )pθ (zs:T ) dzs:T . (2)

The base model is now pθ (zs ) = pθ (zs:T )dzs+1:T , and the super-resolution model is now
R

pθ (x0 |zs ) = p(xT ) Tt=1 pθ (xt−1 |xt , zs )dx1:T , where

R Q

pθ (xt−1 |xt , zs ) = N (xt−1 ; µθ (xt , t, zs , s), Σθ (xt , t, zs , s)).

The reason truncating the low resolution reverse process is a form of data augmentation is
that the training procedure for pθ (x0 |zs ) involves conditioning on noisy zs ∼ q(zs |z0 ), which,
up to scaling, is z0 augmented with Gaussian noise. To be more precise about training a
cascading pipeline with truncated conditioning augmentation, let us examine the ELBO for
psθ (x0 ) in Eq. (2). We can treat psθ (x0 ) as a VAE with a diffusion model prior, a diffusion
model decoder, and the approximate posterior
T
Y
q(x1:T , z1:T |x0 , z0 ) = q(xt |xt−1 )q(zt |zt−1 ),
t=1

7
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

which runs forward processes independently on a low and high resolution pair. The ELBO is
" #
X
s
− log pθ (x0 ) ≤ Eq LT (z0 ) + DKL (q(zt−1 |zt , z0 ) k pθ (zt−1 |zt )) − log pθ (x0 |zs ) ,
t>s

where LT (z0 ) = DKL (q(zT |z0 ) k p(zT )). Note that the sum over t is truncated at s, and the
decoder pθ (x0 |zs ) is the super-resolution model conditioned on zs . The decoder itself has an
ELBO of the form − log pθ (x0 |zs ) ≤ Lθ (x0 |zs ), where
" #
X
Lθ (x0 |zs ) = Eq LT (x0 ) + DKL (q(xt−1 |xt , x0 ) k pθ (xt−1 |xt , zs )) − log pθ (x0 |x1 , zs ) .
t>1

Thus we have an ELBO for the combined model

" #
X
s
− log pθ (x0 ) ≤ Eq LT (z0 ) + DKL (q(zt−1 |zt , z0 ) k pθ (zt−1 |zt )) + Lθ (x0 |zs ) . (3)
t>s

It is apparent that optimizing Eq. (3) trains the low and high resolution models separately.
For a fixed value of s, the low resolution process is trained up to the truncation timestep s,
and the super-resolution model is trained on a conditioning signal corrupted using the low
resolution forward process stopped at timestep s.
In practice, since we pursue sample quality as our main objective, we do not use these
ELBO expressions directly when training models with learnable reverse process variances.
Rather, we train on the “simple” unweighted loss or the hybrid loss described in Section 2,
and the particular loss we use is considered a hyperparameter reported in Appendix B.
We would like to search over multiple values of s to select for the best sample quality. To
make this search practical, we avoid retraining models by amortizing a single super-resolution
model over uniform random s at training time. Because each possible truncation time
corresponds to a distinct super-resolution task, the super-resolution model for µθ and Σθ
must take zs as input along with s, and this can be accomplished using a single network
with an extra time embedding input for s. We leave the low resolution model training
unchanged, because the standard diffusion training procedure already trains with random s.
The complete training procedure for a two-stage cascading pipeline is listed in Algorithm 1.

3.3 Non-truncated Conditioning Augmentation

Another form of conditioning augmentation, which we call non-truncated conditioning augmen-
tation, uses the same model modifications and training procedure as truncated conditioning
augmentation (Section 3.2). The only difference is at sampling time. Instead of truncating the
low resolution reverse process, in non-truncated conditioning augmentation we always sample
z0 using the full, non-truncated low resolution reverse process; then we corrupt z0 using the
forward process into z0s ∼ q(zs |z0 ) and feed the corrupted z0s into the super-resolution model.
The main advantage of non-truncated conditioning augmentation over truncated con-
ditioning augmentation is a practical one during the search phase over s. In the case of
truncated augmentation, if we want to run the super-resolution model over all s in parallel,
we must store all low resolution samples zs for all values of s considered. In the case of

8
 Cascaded Diffusion Models

Algorithm 1 Training a two-stage CDM with Gaussian conditioning augmentation

1: repeat . Train base model
2: (z0 , c) ∼ p(z, c) . Sample low-resolution image and label
3: t ∼ U({1, . . . , T })
4:  ∼ N√(0, I) √
5: zt = ᾱt z0 + 1 − ᾱt 
2
6: θ ← θ − η∇θ kθ (zt , t, c) − k . Simple loss (can be replaced with a hybrid loss)
7: until converged
8: repeat . Train super-resolution model (in parallel with the base model)
9: (x0 , z0 , c) ∼ p(x, z, c) . Sample low- and high-resolution images and label
10: s, t ∼ U({1, . . . , T })
11: z , x√∼ N (0, I)√ . Note: z , x should have the same shapes as z0 , x0 , respectively
12: zt = √ᾱs z0 + √1 − ᾱs z . Apply Gaussian conditioning augmentation
13: xt = ᾱt x0 + 1 − ᾱt x
2
14: θ ← θ − η∇θ kθ (xt , t, zs , s, c) − x k
15: until converged

Algorithm 2 Sampling from a two-stage CDM with Gaussian conditioning augmentation

Require: c: class label
Require: s: conditioning augmentation truncation time
1: zT ∼ N (0, I)
2: if using truncated conditioning augmentation then
3: for t = T, . . . , s + 1 do
4: zt−1 ∼ pθ (zt−1 |zt , c)
5: end for
6: else
7: for t = T, . . . , 1 do
8: zt−1 ∼ pθ (zt−1 |zt , c)
9: end for
10: zs ∼ q(zs |z0 ) . Overwrite previously sampled value of zs
11: end if
12: xT ∼ N (0, I)
13: for t = T, . . . , 1 do
14: xt−1 ∼ pθ (xt−1 |xt , zs , c)
15: end for
16: return x0

non-truncated augmentation, we need to store the low resolution samples just once, since
sampling z0s ∼ q(zs |z0 ) is computationally inexpensive. These sampling procedures are listed
in Algorithm 2.

Truncated and non-truncated conditioning augmentation should perform similarly be-

cause zs and z0s should have similar marginal distributions if the low resolution model is
trained well enough. Indeed, in Section 4.3, we empirically find that sample quality metrics
are similar for both truncated and non-truncated conditioning augmentation.

9
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

Figure 5: Classwise Synthetic 256×256 ImageNet images. Each row represents a specific
ImageNet class. Classes from top to bottom - Flamingo (130), White Wolf (270),
Tiger (292), Monarch Butterfly (323), Zebra (340) and Dome (538).

10
 Cascaded Diffusion Models

Figure 6: Classwise Synthetic 256×256 ImageNet images. Each row represents a specific
ImageNet class. Classes from top to bottom - Greenhouse (580), Model T (661),
Streetcar (829), Comic Book (917), Crossword Puzzle (918), Cheeseburger (933).

11
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

4. Experiments
We designed experiments to improve the sample quality metrics of cascaded diffusion models
on class-conditional ImageNet generation. Our cascading pipelines consist of class-conditional
diffusion models at all resolutions, so class information is injected at all resolutions: see Fig. 4.
Our final ImageNet results are described in Section 4.1.
To give insight into our cascading pipelines, we begin with improvements on a baseline
non-cascaded model at the 64×64 resolution (Section 4.2), then we show that cascading up
to 64×64 improves upon our best non-cascaded 64×64 model, but only in conjunction with
conditioning augmentation. We also show that truncated and non-truncated conditioning
augmentation perform equally well (Section 4.3), and we study random Gaussian blur augmen-
tation to train super-resolution models to resolutions of 128×128 and 256×256 (Section 4.4).
Finally, we verify that conditioning augmentation is also effective on the LSUN dataset (Yu
et al., 2015) and therefore is not specific to ImageNet (Section 4.5).
We cropped and resized the ImageNet dataset (Russakovsky et al., 2015) in the same
manner as BigGAN (Brock et al., 2019). We report Inception scores using the standard
practice of generating 50k samples and calculating the mean and standard deviation over 10
splits (Salimans et al., 2016). Generally, throughout our experiments, we selected models and
performed early stopping based on FID score calculated over 10k samples, but all reported
FID scores are calculated over 50k samples for comparison with other work (Heusel et al.,
2017). The FID scores we used for model selection and reporting model performance are
calculated against training set statistics according to common practice, but since this can be
seen as overfitting on the performance metric, we additionally report model performance
using FID scores calculated against validation set statistics. We also report results on
Classification Accuracy Score (CAS), which was proposed by Ravuri and Vinyals (2019) due
to their findings that non-GAN models may score poorly on FID and IS despite generating
visually appealing samples and that FID and IS are not correlated (sometimes anti-correlated)
with performance on downstream tasks.

4.1 Main Cascading Pipeline Results

Table 1a reports the main results on the cascaded diffusion model (CDM ), for the 64×64,
128×128, and 256×256 ImageNet dataset resolutions, along with baselines. CDM outperforms
BigGAN-deep in terms of FID score on the image resolutions considered, but GANs perform
better in terms of Inception score when their truncation parameter is optimized for Inception
score (Brock et al., 2019). We also outperform concurrently released diffusion models that
do not use classifier guidance to boost sample quality scores (Dhariwal and Nichol, 2021).
See Fig. 8 for a qualitative assessment of sample quality and diversity compared to VQ-VAE-
2 (Razavi et al., 2019) and BigGAN-deep (Brock et al., 2019), and see Figs. 5 and 6 for
examples of generated images.
Table 1b reports the results on Classification Accuracy Score (CAS) (Ravuri and Vinyals,
2019) for our models at the 128×128 and 256×256 resolutions. We find that CDM outperforms
VQ-VAE-2 and BigGAN-deep at both resolutions by a significant margin on the CAS metric,
suggesting better potential performance on downstream tasks. Figure 7 compares class-wise
classification accuracy scores between classifiers trained on real training data, and CDM
samples. The CDM classifier outperforms real data on 96 classes compared to 6 and 31

12
 Cascaded Diffusion Models

classes by BigGAN-deep and VQ-VAE-2 respectively. We also show samples from classes
with best and worst accuracy scores in Appendix Figure 11 and 12.

Our cascading pipelines are structured as a 32×32 base model, a 32×32→64×64 super-
resolution model, followed by 64×64→128×128 or 64×64→256×256 super-resolution models.
Models at 32×32 and 64×64 resolutions use 4000 diffusion timesteps and architectures similar
to DDPM (Ho et al., 2020) and Improved DDPM (Nichol and Dhariwal, 2021). Models
at 128×128 and 256×256 resolutions use 100 sampling steps, determined by post-training
hyperparameter search (Section 4.4), and they use architectures similar to SR3 (Saharia
et al., 2021). All base resolution and super-resolution models are conditioned on class labels.
See Appendix B for details.

1
Classification Accuracy

0.8

0.6

0.4

0.2
Real Training Data
0 CDM Samples
0 200 400 600 800 1000
Class ID

Figure 7: Classwise Classification Accuracy Score comparison between real data (blue) and
generated data (red) at the 256×256 resolution. Accompanies Table 1b.

4.2 Baseline Model Improvements

To set a strong baseline for class-conditional ImageNet generation at the 64×64 resolution,
we reproduced and improved upon a 4000 timestep non-cascaded 64×64 class-conditional
diffusion model from Improved DDPM (Nichol and Dhariwal, 2021). Our reimplementation
used dropout and was trained longer than reported by Nichol and Dhariwal; we found that
adding dropout generally slowed down convergence of FID and Inception scores, but improved
their best values over the course of a longer training period. We further improved the training
set FID score and Inception score by adding noise to the trained model’s samples using the
forward process to the 2000 timestep point, then restarting the reverse process from that
point. See Table 2a for the resulting sample quality metrics.

13
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

FID FID
Model IS
vs train vs validation
32×32 resolution
CDM (ours) 1.11 1.99 26.01 ± 0.59
64×64 resolution
BigGAN-deep, by (Dhariwal and Nichol, 2021) 4.06
Improved DDPM (Nichol and Dhariwal, 2021) 2.92
ADM (Dhariwal and Nichol, 2021) 2.07
CDM (ours) 1.48 2.48 67.95 ± 1.97
128×128 resolution
BigGAN-deep (Brock et al., 2019) 5.7 124.5
BigGAN-deep, max IS (Brock et al., 2019) 25 253
LOGAN (Wu et al., 2019) 3.36 148.2
ADM (Dhariwal and Nichol, 2021) 5.91
CDM (ours) 3.52 3.76 128.80 ± 2.51
256×256 resolution
BigGAN-deep (Brock et al., 2019) 6.9 171.4
BigGAN-deep, max IS (Brock et al., 2019) 27 317
VQ-VAE-2 (Razavi et al., 2019) 31.11
Improved DDPM (Nichol and Dhariwal, 2021) 12.26
SR3 (Saharia et al., 2021) 11.30
ADM (Dhariwal and Nichol, 2021) 10.94 100.98
ADM+upsampling (Dhariwal and Nichol, 2021) 7.49 127.49
CDM (ours) 4.88 4.63 158.71 ± 2.26

(a) Class-conditional ImageNet sample quality results for classifier guidance-free methods

Model Top-1 Accuracy Top-5 Accuracy

128×128 resolution
Real 68.82% 88.79%
BigGAN-deep (Brock et al., 2019) 40.64% 64.44%
HAM (De Fauw et al., 2019) 54.05% 77.33%
CDM (ours) 59.84% 81.79%
256×256 resolution
Real 73.09% 91.47%
BigGAN-deep (Brock et al., 2019) 42.65% 65.92%
VQ-VAE-2 (Razavi et al., 2019) 54.83% 77.59%
CDM (ours) 63.02% 84.06%

(b) Classification Accuracy Score (CAS) results

Table 1: Main results. Numbers are bolded only when at least two are available for compari-
son. CAS for real data and other models are from Ravuri and Vinyals (2019).

14
 Cascaded Diffusion Models

CDM (ours) VQ-VAE-2 BigGAN-deep

Figure 8: Comparing the quality and diversity of model samples in selected 256×256 Ima-
geNet classes {Tench(0), Goldfish(1) and Ostrich(9)}. VQVAE-2 and BigGAN
samples are taken from Razavi et al. (2019).

15
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

FID FID
Model IS
vs train vs validation
Improved DDPM (Nichol and Dhariwal, 2021) 2.92
Our reimplementation 2.44 2.91 49.81 ± 0.65
+ more sampling steps 2.35 2.91 52.72 ± 1.15

(a) Improvements to a non-cascaded baseline

FID FID
Conditioning IS
vs train vs validation
No cascading 2.35 2.91 52.72 ± 1.15
16×16→64×64 cascading
s=0 6.02 5.84 35.59 ± 1.19
s = 101 3.41 3.67 44.72 ± 1.12
s = 1001 2.13 2.79 54.47 ± 1.05

(b) Small-scale ablation comparing no cascading to 16×16→64×64 cascading

Table 2: 64×64 ImageNet sample quality: ablations.

4.3 Conditioning Augmentation Experiments up to 64×64

Building on our reimplementation in Section 4.2, we verify in a small scale experiment that
cascading improves sample quality at the 64×64 resolution. We train a two-stage cascading
pipeline that comprises a 16×16 base model and a 16×16→64×64 super-resolution model.
The super-resolution model architecture is identical to the best 64×64 non-cascaded baseline
model in Section 4.2, except for the trivial modification of adding in the low resolution image
conditioning information by channelwise concatenation at the input (see Section 2).
See Table 2b and Fig. 9 for the results of this 16×16→64×64 cascading pipeline. Interest-
ingly, we find that without conditioning augmentation, the cascading pipeline attains lower
sample quality than the non-cascaded baseline 64×64 model; the FID score, for example,
degrades from 2.35 to 6.02. With sufficient conditioning augmentation, however, the sample
quality of the cascading pipeline becomes better than the non-cascaded baseline. We train
two super-resolution models with non-truncated conditioning augmentation, one at trun-
cation time s = 101 and another at s = 1001 (we could have amortized both into a single
model, but we chose not to do so in this experiment to prevent potential model capacity
issues from confounding the results). The first model achieves better sample quality than
the non-augmented model but is still worse than the non-cascaded baseline. The second
model achieves a FID score of 2.13, outperforming the non-cascaded baseline. Conditioning
augmentation is therefore crucial to improve sample quality in this particular cascading
pipeline.
To further improve sample quality at the 64×64 resolution, we found it helpful to increase
model sizes and to switch to a cascading pipeline starting with a 32×32 base resolution
model. We train a 32×32 base model applying random left-right flips, which we found to
help 32×32 scores at the expense of longer training times. Training without random flips, the

16
 Cascaded Diffusion Models

(a) 16×16 base (b) 16×16→64×64 super-resolution, s = 0

(c) 16×16→64×64 super-resolution, s = 101 (d) 16×16→64×64 super-resolution, s = 1001

Figure 9: Generated images for varying amounts of conditioning augmentation (non-

truncated) in a small-scale 16×16→64×64 pipeline for ablation purposes. Ac-
companies Table 2b.

17
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

best 32×32 resolution FID score is 1.25 at 300k training steps, while training with random
flips it is 1.11 at 700k training steps. The 32×32→64×64 super-resolution model is now
amortized over the truncation time s by providing s as an extra time embedding input to
the network (Section 2), allowing us to perform a more fine grained search over s without
retraining the model.
Table 3a displays the resulting sample quality scores for both truncated and non-truncated
augmentation. The sample quality metrics improve and then degrade non-monotonically
as the truncation time is increased. This indicates that moderate amounts of conditioning
augmentation are beneficial to sample quality of the cascading pipeline, but too much
conditioning augmentation causes the super-resolution model to behave as a non-conditioned
model unable to benefit from cascading. For comparison, Table 3b shows sample quality when
the super-resolution model is conditioned on ground truth data instead of generated data.
Here, sample quality monotonically degrades as truncation time is increased. Conditioning
augmentation is therefore useful precisely when conditioning on generated samples, so as a
technique it is uniquely suited to cascading pipelines.
Based on these findings on non-monotonicity of sample quality with respect to truncation
time, we conclude that conditioning augmentation works because it alleviates compounding
error from a train-test mismatch for the super-resolution model. This occurs when low-
resolution model samples are out of distribution compared to the ground truth data on
which the super-resolution model is trained. A sufficient amount of Gaussian conditioning
augmentation prevents the super-resolution model from attempting to upsample erroneous,
out-of-distribution details in the low resolution generated samples. In contrast, sample
quality degrades monotonically with respect to truncation time when conditioning the
super-resolution model on ground truth data, because there is no such train-test mismatch.
Table 3a additionally shows that truncated and non-truncated conditioning augmentation
are approximately equally effective at improving sample quality of the cascading pipeline,
albeit at different values of the truncation time parameter. Thus we generally recommend
non-truncated augmentation due to its practical benefits described in Section 3.3.

4.4 Experiments at 128×128 and 256×256

While we found Gaussian noise augmentation to be a key ingredient to boost the performance
of our cascaded models at low resolutions, our initial experiments with similar augmentations
for 128×128 and 256×256 upsampling yielded negative results. Hence, we explore Gaussian
blurring augmentation for these resolutions. As mentioned in Section 3.1, we apply the
blurring augmentation 50% of the time during training, and use no blurring during inference.
We explored other settings (e.g. applying blurring to all training examples, and using varying
amounts of blurring during inference), but found this to be most effective in our initial
experiments.
Table 4a shows the results of applying Gaussian blur augmentation to the 64×64 →
256×256 super-resolution model. While any amount of blurring helps improve the scores
of the 256×256 samples over the baseline model with no blur, we found that sampling
σ ∼ U(0.4, 0.6) gives the best results. Table 4b shows further improvements from class
conditioning, large batch training, and random flip augmentation for the super-resolution
model. While we find class conditioning helpful for upsampling at low resolution settings, it is

18
 Cascaded Diffusion Models

FID FID
Conditioning IS
vs train vs validation
No conditioning augmentation (baseline)
s=0 1.71 2.46 61.34 ± 1.58
Truncated conditioning augmentation
s = 251 1.50 2.44 66.76 ± 1.76
s = 501 1.48 2.48 67.95 ± 1.97
s = 751 1.48 2.51 68.48 ± 1.77
s = 1001 1.49 2.51 67.95 ± 1.51
s = 1251 1.51 2.54 67.20 ± 1.94
s = 1501 1.54 2.56 67.09 ± 1.67
Non-truncated conditioning augmentation
s = 251 1.58 2.50 66.21 ± 1.51
s = 501 1.53 2.51 67.59 ± 1.85
s = 751 1.48 2.47 67.48 ± 1.31
s = 1001 1.49 2.48 66.51 ± 1.59
s = 1251 1.48 2.46 66.28 ± 1.49
s = 1501 1.50 2.47 65.59 ± 0.86

(a) Base model for low resolution conditioning

FID FID
Conditioning IS
vs train vs validation
Ground truth training data
s=0 0.76 1.76 74.84 ± 1.43
s = 251 0.87 1.85 71.79 ± 0.89
s = 501 0.92 1.91 70.68 ± 1.26
s = 751 0.95 1.94 69.93 ± 1.40
s = 1001 0.98 1.97 69.03 ± 1.26
s = 1251 1.03 1.99 67.92 ± 1.65
s = 1501 1.11 2.04 66.7 ± 1.21
Ground truth validation data
s=0 1.20 0.59 64.33 ± 1.24
s = 251 1.27 0.96 63.17 ± 1.19
s = 501 1.32 1.17 62.65 ± 0.76
s = 751 1.38 1.32 62.21 ± 0.94
s = 1001 1.42 1.44 61.53 ± 1.39
s = 1251 1.47 1.54 60.58 ± 0.93
s = 1501 1.53 1.64 60.02 ± 0.84

(b) Ground truth for low resolution conditioning

Table 3: 64×64 ImageNet sample quality: large scale experiment comparing truncated
and non-truncated conditioning augmentation for 32×32→64×64 cascading, using
amortized truncation time conditioning.

19
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

5.6

5.3

FID
4.9

FID vs Train
4.6 FID vs Validation
4 8 16 25 100 1000
Inference Steps

Figure 10: FID on 256×256 images vs inference steps in 64×64 → 256×256 super-resolution.

FID FID
Blur σ IS
vs train vs validation
σ = 0 (no blur) 7.26 6.42 134.53 ± 2.97
σ ∼ U(0.4, 0.6) 6.18 5.57 142.71 ± 2.83
σ ∼ U(0.4, 0.8) 6.90 6.31 136.57 ± 4.34
σ ∼ U(0.4, 1.0) 6.35 5.76 141.40 ± 4.34

(a) Gaussian blur noise in conditioning

FID FID
Model IS
vs train vs validation
Baseline 6.18 5.57 142.71 ± 2.83
+ Class Conditioning 5.75 5.27 152.17 ± 2.29
+ Large Batch Training 5.00 4.71 157.84 ± 2.60
+ Flip LR 4.88 4.63 158.71 ± 2.26

(b) Further improvements on super-resolution

Table 4: 256×256 ImageNet sample quality: experiments on 64×64 → 256×256 super-

resolution.

20
 Cascaded Diffusion Models

interesting that it still gives a huge boost to the upsampling performance at high resolutions
even when the low resolution inputs at 64×64 can be sufficiently informative. We also
found increasing the training batch size from 256 to 1024 further improved performance by a
significant margin. We also obtain marginal improvements by training the super-resolution
model on randomly flipped data.
Since the sampling cost increases quadratically with the target image resolution, we
attempt to minimize the number of denoising iterations for our 64×64 → 256×256 and
64×64 → 128×128 super-resolution models. To this end, we train these super-resolution
models with continuous noise conditioning, like Saharia et al. (2021) and Chen et al. (2021),
and tune the noise schedule for a given number of steps during inference. This tuning is
relatively inexpensive as we do not need to retrain the models. We report all results using
100 inference steps for these models. Figure 10 shows FID vs number of inference steps
for our 64×64 → 256×256 model. The FID score deteriorates marginally even when using
just 4 inference steps. Interestingly, we do not observe any concrete improvement in FID by
increasing the number of inference steps from 100 to 1000.

4.5 Experiments on LSUN

While the main results of this work are on class-conditional ImageNet generation, here we
study the effectiveness of non-truncated conditioning augmentation for a 64×64→128×128
cascading pipeline on the LSUN Bedroom and Church datasets (Yu et al., 2015) in order to
verify that conditioning augmentation is not an ImageNet-specific method. LSUN Bedroom
and Church are two separate unconditional datasets that do not have any class labels,
so our study here additionally verifies the effectiveness of conditioning augmentation for
unconditional generation.
Table 5 displays our LSUN sample quality results, which confirm that a nonzero amount
of conditioning augmentation is beneficial to sample quality. (The relatively large FID
scores between generated examples and the validation sets are explained by the fact that
the LSUN Church and Bedroom validation sets are extremely small, consisting of only 300
examples each.) We observe a similar effect as our ImageNet results in Table 3b: because the
super-resolution model is conditioned on base model samples, the sample quality improves
then degrades non-monotonically as the truncation time s is increased. See Appendix A for
examples of images generated by our LSUN models.

5. Related Work
One way to formulate cascaded diffusion models is to modify the original diffusion formalism
of a forward process q(x0:T ) at single resolution so that the transition q(xt |xt−1 ) performs
downsampling at certain intermediate timesteps, for example at t ∈ S := {T /4, 2T /4, 3T /4}.
The reverse process would then be required to perform upsampling at those timesteps, similar
to our cascaded models here. However, there is no guarantee that the reverse transitions
at the timesteps in S are conditional Gaussian, unlike the guarantee for reverse transitions
at other timesteps for sufficiently slow diffusion. By contrast, our cascaded diffusion model
construction dedicates entire conditional diffusion models for these upsampling steps, so it is
specified more flexibly.

21
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

FID FID
Conditioning
vs train vs validation
LSUN Bedroom
s=0 2.30 40.68
s = 251 2.06 40.47
s = 501 2.08 40.44
s = 751 2.14 40.45
s = 1001 2.18 40.53
s = 1251 2.24 40.58
s = 1501 2.28 40.58
LSUN Church
s=0 3.29 42.21
s = 251 2.97 42.14
s = 501 2.93 42.17
s = 751 2.89 42.20
s = 1001 2.86 42.26
s = 1251 2.83 42.28
s = 1501 2.84 42.31

Table 5: 128×128 LSUN sample quality: non-truncated conditioning augmentation for

a 64×64→128×128 cascading pipeline using the base model for low resolution
conditioning.

Recent interest in diffusion models (Sohl-Dickstein et al., 2015) started with work con-
necting diffusion models to denoising score matching over multiple noise scales (Ho et al.,
2020; Song and Ermon, 2019). There have been a number of improvements and alternatives
proposed to the diffusion framework, for example generalization to continuous time (Song
et al., 2021b), deterministic sampling (Song et al., 2021a), adversarial training (Jolicoeur-
Martineau et al., 2021), and others (Gao et al., 2021). For simplicity, we base our models on
DDPM (Ho et al., 2020) with modifications from Improved DDPM (Nichol and Dhariwal,
2021) to stay close to the original diffusion framework.
Cascading pipelines have been investigated in work on VQ-VAEs (van den Oord et al.,
2016c; Razavi et al., 2019) and autoregressive models (Menick and Kalchbrenner, 2019).
Cascading pipelines have also been investigated for diffusion models, such as SR3 (Saharia
et al., 2021), Improved DDPM (Nichol and Dhariwal, 2021), and concurrently in ADM (Dhari-
wal and Nichol, 2021). Our work here focuses on improving cascaded diffusion models for
ImageNet generation and is distinguished by the extensive study on conditioning augmenta-
tion and deeper cascading pipelines. Our conditioning augmentation work also resembles
scheduled sampling in autoregressive sequence generation (Bengio et al., 2015), where noise
is used to alleviate the mismatch between train and inference conditions.
Concurrent work (Dhariwal and Nichol, 2021) showed that diffusion models are capable
of generating high quality ImageNet samples using an improved architecture, named ADM,
and a classifier guidance technique in which a class-conditional diffusion model sampler is
modified to simultaneously take gradient steps to maximize the score of an extra trained image

22
 Cascaded Diffusion Models

classifier. By contrast, our work focuses solely on improving sample quality by cascading, so
we avoid introducing extra model elements such as the image classifier. We are interested in
avoiding classifier guidance because the FID and Inception score sample quality metrics that
we use to evaluate our models are themselves computed on activations of an image classifier
trained on ImageNet, and therefore classifier guidance runs the risk of cheating these metrics.
Avoiding classifier guidance comes at the expense of using thousands of diffusion timesteps
in our low resolution models, where ADM uses hundreds. ADM with classifier guidance
outperforms our models in terms of FID and Inception scores, while our models outperform
ADM without classifier guidance as reported by Dhariwal and Nichol. Our work is a showcase
of the effectiveness of cascading alone in a pure generative model, and since classifier guidance
and cascading complement each other as techniques to improve sample quality and can be
applied together, we expect classifier guidance would improve our results too.

6. Conclusion
We have shown that cascaded diffusion models are capable of outperforming state-of-the-art
generative models on the ImageNet class-conditional generation benchmark when paired
with conditioning augmentation, our technique of introducing data augmentation into the
conditioning information of super-resolution models. Our models outperform BigGAN-deep
and VQ-VAE-2 as measured by FID score and classification accuracy score. We found that
conditioning augmentation helps sample quality because it combats compounding error in
cascading pipelines due to train-test mismatch in super-resolution models, and we proposed
practical methods to train and test models amortized over varying levels of conditioning
augmentation.
Although there could be negative impact of our work in the form of malicious uses of
image generation, our work has the potential to improve beneficial downstream applications
such as data compression while advancing the state of knowledge in fundamental machine
learning problems. We see our results as a conceptual study of the image synthesis capabilities
of diffusion models in their original form with minimal extra techniques, and we hope our
work serves as inspiration for future advances in the capabilities of diffusion models.

Acknowledgments

We thank Jascha Sohl-Dickstein, Douglas Eck and the Google Brain team for feedback,
research discussions and technical assistance.

23
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

Appendix A. Samples

Figure 11: Samples from classes with best relative classification accuracy score. Each row represents
a specific ImageNet class. Classes from top to bottom - Tiger Cat (282), Gong (577),
Coffee Mug (504), Squirrel Monkey (382), Miniature Schnauzer (196) and Corn (987).

24
 Cascaded Diffusion Models

Figure 12: Samples from classes with worst relative classification accuracy score. Each row represents
a specific ImageNet class. Classes from top to bottom - Letter Opener (623), Plate (923),
Overskirt (689), Tobacco Shop (860), Black-and-tan Coonhound (165) and Bathtub
(435).

25
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

Figure 13: Samples from LSUN 128x128: bedroom subset (first six rows) and church subset (last
six rows).

26
 Cascaded Diffusion Models

Appendix B. Hyperparameters
B.1 ImageNet
Here we give the hyperparameters of the models in our ImageNet cascading pipelines. Each
model in the pipeline is described by its diffusion process, its neural network architecture, and
its training hyperparameters. Architecture hyperparameters, such as the base channel count
and the list of channel multipliers per resolution, refer to hyperparameters of the U-Net in
DDPM and related models (Ho et al., 2020; Nichol and Dhariwal, 2021; Saharia et al., 2021;
Salimans et al., 2017). The cosine noise schedule and the hybrid loss method of learning re-
verse process variances are from Improved DDPM (Nichol and Dhariwal, 2021). Some models
are conditioned on ᾱt for post-training sampler tuning (Chen et al., 2021; Saharia et al., 2021).

32×32 base model

• Architecture • Diffusion
– Base channels: 256 – Timesteps: 4000
– Channel multipliers: 1, 2, 3, 4 – Noise schedule: cosine
– Residual blocks per resolution: 6 – Learned variances: yes
– Attention resolutions: 8, 16 – Loss: hybrid
– Attention heads: 4
• Training
– Optimizer: Adam
– Batch size: 2048
– Learning rate: 1e-4
– Steps: 700000
– Dropout: 0.1
– EMA: 0.9999
– Hardware: 256 TPU-v3 cores

27
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

32×32→64×64 super-resolution

• Architecture • Diffusion
– Base channels: 256 – Timesteps: 4000
– Channel multipliers: 1, 2, 3, 4 – Noise schedule: cosine
– Residual blocks per resolution: 5 – Learned variances: yes
– Attention resolutions: 8, 16 – Loss: hybrid
– Attention heads: 4
• Training
– Optimizer: Adam
– Batch size: 2048
– Learning rate: 1e-4
– Steps: 400000
– Dropout: 0.1
– EMA: 0.9999
– Hardware: 256 TPU-v3 cores

64×64→128×128 super-resolution

• Architecture • Diffusion (Training)

– Base channels: 128 – Timesteps: 2000
– Channel multipliers: 1, 2, 4, 8, 8 – Noise schedule: linear
– Residual blocks per resolution: 3 – Learned variances: no
– Attention resolutions: 16 – Loss: simple
– Attention heads: 1 – Continuous noise conditioning
• Training • Diffusion (Inference)
– Optimizer: Adam – Timesteps: 100
– Batch size: 1024 – Noise schedule: linear
– Learning rate: 1e-4
– Steps: 500000
– Dropout: 0.0
– EMA: 0.9999
– Hardware: 128 TPU-v3 cores

28
 Cascaded Diffusion Models

64×64→256×256 super-resolution

• Architecture • Diffusion (Training)

– Base channels: 128 – Timesteps: 2000
– Channel multipliers: 1, 2, 4, 4, 8, 8 – Noise schedule: linear
– Residual blocks per resolution: 3 – Learned variances: no
– Attention resolutions: 16 – Loss: simple
– Attention heads: 1 – Continuous noise conditioning
• Training • Diffusion (Inference)
– Optimizer: Adam – Timesteps: 100
– Batch size: 1024 – Noise schedule: linear
– Learning rate: 1e-4
– Steps: 500000
– Dropout: 0.0
– EMA: 0.9999
– Hardware: 128 TPU-v3 cores

29
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

B.2 LSUN
Here we give the hyperparameters of our LSUN Bedroom and Church pipelines. We used
the same hyperparameters for both datasets.
64×64 base model

• Architecture • Diffusion (Training)

– Base channels: 128 – Noise schedule: cosine
– Channel multipliers: 1, 2, 3, 4 – Learned variances: no
– Residual blocks per resolution: 3 – Loss: simple
– Attention resolutions: 8, 16, 32 – Continuous noise conditioning
– Attention heads dimension: 64 • Diffusion (Inference)
• Training – Timesteps: 256
– Optimizer: Adam – Noise schedule: cosine
– Batch size: 2048
– Learning rate: 3e-4
– Steps: 100000
– Dropout: 0.1
– EMA: 0.9999
– Hardware: 64 TPU-v3 cores

64×64→128×128 super-resolution

• Architecture • Diffusion (Training)

– Base channels: 64 – Noise schedule: cosine
– Channel multipliers: 1, 2, 4, 6, 8 – Learned variances: no
– Residual blocks per resolution: 3 – Loss: simple
– Attention resolutions: 8, 16, 32 – Continuous noise conditioning
– Attention heads dimension: 64 • Diffusion (Inference)
• Training – Timesteps: 256
– Optimizer: Adam – Noise schedule: cosine
– Batch size: 1024
– Learning rate: 2e-4
– Steps: 220000
– Dropout: 0.1
– EMA: 0.9999
– Hardware: 64 TPU-v3 cores

30
 Cascaded Diffusion Models

References
S. Bengio, O. Vinyals, N. Jaitly, and N. Shazeer. Scheduled sampling for sequence prediction
with recurrent neural networks. Advances in Neural Information Processing Systems, 2015.

A. Brock, J. Donahue, and K. Simonyan. Large scale GAN training for high fidelity natural
image synthesis. In International Conference on Learning Representations, 2019.

N. Chen, Y. Zhang, H. Zen, R. J. Weiss, M. Norouzi, and W. Chan. WaveGrad: Estimating

gradients for waveform generation. International Conference on Learning Representations,
2021.

J. De Fauw, S. Dieleman, and K. Simonyan. Hierarchical autoregressive image models with

auxiliary decoders. arXiv preprint arXiv:1903.04933, 2019.

P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis. arXiv preprint
arXiv:2105.05233, 2021.

L. Dinh, J. Sohl-Dickstein, and S. Bengio. Density estimation using Real NVP. International
Conference on Learning, 2017.

R. Gao, Y. Song, B. Poole, Y. N. Wu, and D. P. Kingma. Learning energy-based models by

diffusion recovery likelihood. International Conference on Learning Representations, 2021.

I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville,

and Y. Bengio. Generative adversarial nets. In Advances in Neural Information Processing
Systems, pages 2672–2680, 2014.

M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter. GANs trained by a

two time-scale update rule converge to a local Nash equilibrium. In Advances in Neural
Information Processing Systems, pages 6626–6637, 2017.

J. Ho, X. Chen, A. Srinivas, Y. Duan, and P. Abbeel. Flow++: Improving flow-based

generative models with variational dequantization and architecture design. In International
Conference on Machine Learning, 2019.

J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In Advances in

Neural Information Processing Systems, pages 6840–6851, 2020.

A. Jolicoeur-Martineau, R. Piché-Taillefer, R. T. d. Combes, and I. Mitliagkas. Adversarial

score matching and improved sampling for image generation. International Conference on
Learning Representations, 2021.

D. P. Kingma and P. Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. In
Advances in Neural Information Processing Systems, pages 10215–10224, 2018.

D. P. Kingma and M. Welling. Auto-encoding variational Bayes. International Conference

on Learning Representations, 2014.

Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro. DiffWave: A Versatile Diffusion

Model for Audio Synthesis. International Conference on Learning Representations, 2021.

31
 Ho, Saharia, Chan, Fleet, Norouzi and Salimans

J. Menick and N. Kalchbrenner. Generating high fidelity images with subscale pixel networks
and multidimensional upscaling. In International Conference on Learning Representations,
2019.

A. Nichol and P. Dhariwal. Improved denoising diffusion probabilistic models. International

Conference on Machine Learning, 2021.

M. Ranzato, S. Chopra, M. Auli, and W. Zaremba. Sequence level training with recurrent
neural networks. International Conference on Learning Representations, 2016.

S. Ravuri and O. Vinyals. Classification Accuracy Score for Conditional Generative Models.
In Advances in Neural Information Processing Systems, volume 32, 2019.

A. Razavi, A. van den Oord, and O. Vinyals. Generating diverse high-fidelity images with
VQ-VAE-2. In Advances in Neural Information Processing Systems, pages 14837–14847,
2019.

O. Ronneberger, P. Fischer, and T. Brox. U-Net: Convolutional networks for biomedical

image segmentation. In International Conference on Medical Image Computing and
Computer-Assisted Intervention, pages 234–241. Springer, 2015.

O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpa-

thy, A. Khosla, M. Bernstein, et al. ImageNet large scale visual recognition challenge.
International Journal of Computer Vision, 115(3):211–252, 2015.

C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi. Image super-resolution

via iterative refinement. arXiv preprint arXiv:2104.07636, 2021.

T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen. Improved

techniques for training gans. In Advances in Neural Information Processing Systems, pages
2234–2242, 2016.

T. Salimans, A. Karpathy, X. Chen, and D. P. Kingma. PixelCNN++: Improving the Pixel-

CNN with discretized logistic mixture likelihood and other modifications. In International
Conference on Learning Representations, 2017.

J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli. Deep unsupervised

learning using nonequilibrium thermodynamics. In International Conference on Machine
Learning, pages 2256–2265, 2015.

J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. International

Conference on Learning Representations, 2021a.

Y. Song and S. Ermon. Generative modeling by estimating gradients of the data distribution.
In Advances in Neural Information Processing Systems, pages 11895–11907, 2019.

Y. Song and S. Ermon. Improved techniques for training score-based generative. Advances
in Neural Information Processing Systems, 2020.

32
 Cascaded Diffusion Models

Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based

generative modeling through stochastic differential equations. International Conference on
Learning Representations, 2021b.

A. van den Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A. Graves, N. Kalchbrenner,

A. Senior, and K. Kavukcuoglu. WaveNet: A generative model for raw audio. arXiv
preprint arXiv:1609.03499, 2016a.

A. van den Oord, N. Kalchbrenner, and K. Kavukcuoglu. Pixel recurrent neural networks.
International Conference on Machine Learning, 2016b.

A. van den Oord, N. Kalchbrenner, O. Vinyals, L. Espeholt, A. Graves, and K. Kavukcuoglu.

Conditional image generation with PixelCNN decoders. In Advances in Neural Information
Processing Systems, pages 4790–4798, 2016c.

A. van den Oord, O. Vinyals, and K. Kavukcuoglu. Neural discrete representation learning.
Advances in Neural Information Processing Systems, 2017.

Y. Wu, J. Donahue, D. Balduzzi, K. Simonyan, and T. Lillicrap. Logan: Latent optimisation

for generative adversarial networks. arXiv preprint arXiv:1912.00953, 2019.

F. Yu, Y. Zhang, S. Song, A. Seff, and J. Xiao. LSUN: Construction of a large-scale image
dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365,
2015.

33