original_content.txt

High-Resolution Image Synthesis with Latent Diffusion Models
Robin Rombach1* Andreas Blattmann1Dominik Lorenz1Patrick Esser
 Bj¨orn Ommer1
1Ludwig Maximilian University of Munich & IWR, Heidelberg University, Germany
 Runway ML
https://github.com/CompVis/latent-diffusion
Abstract
By decomposing the image formation process into a se-
quential application of denoising autoencoders, diffusion
models (DMs) achieve state-of-the-art synthesis results on
image data and beyond. Additionally, their formulation al-
lows for a guiding mechanism to control the image gen-
eration process without retraining. However, since these
models typically operate directly in pixel space, optimiza-
tion of powerful DMs often consumes hundreds of GPU
days and inference is expensive due to sequential evalu-
ations. To enable DM training on limited computational
resources while retaining their quality and flexibility, we
apply them in the latent space of powerful pretrained au-
toencoders. In contrast to previous work, training diffusion
models on such a representation allows for the first time
to reach a near-optimal point between complexity reduc-
tion and detail preservation, greatly boosting visual fidelity.
By introducing cross-attention layers into the model archi-
tecture, we turn diffusion models into powerful and flexi-
ble generators for general conditioning inputs such as text
or bounding boxes and high-resolution synthesis becomes
possible in a convolutional manner. Our latent diffusion
models (LDMs) achieve new state-of-the-art scores for im-
age inpainting and class-conditional image synthesis and
highly competitive performance on various tasks, includ-
ing text-to-image synthesis, unconditional image generation
and super-resolution, while significantly reducing computa-
tional requirements compared to pixel-based DMs.
1. Introduction
Image synthesis is one of the computer vision fields with
the most spectacular recent development, but also among
those with the greatest computational demands. Espe-
cially high-resolution synthesis of complex, natural scenes
is presently dominated by scaling up likelihood-based mod-
els, potentially containing billions of parameters in autore-
gressive (AR) transformers [66,67]. In contrast, the promis-
ing results of GANs [3, 27, 40] have been revealed to be
mostly confined to data with comparably limited variability
as their adversarial learning procedure does not easily scale
to modeling complex, multi-modal distributions. Recently,
diffusion models [82], which are built from a hierarchy of
denoising autoencoders, have shown to achieve impressive
*The first two authors contributed equally to this work.Inputours ( f= 4)
PSNR: 27:4R-FID: 0:58DALL-E ( f= 8)
PSNR: 22:8R-FID: 32:01VQGAN ( f= 16 )
PSNR: 19:9R-FID: 4:98
Figure 1. Boosting the upper bound on achievable quality with
less agressive downsampling. Since diffusion models offer excel-
lent inductive biases for spatial data, we do not need the heavy spa-
tial downsampling of related generative models in latent space, but
can still greatly reduce the dimensionality of the data via suitable
autoencoding models, see Sec. 3. Images are from the DIV2K [1]
validation set, evaluated at 5122px. We denote the spatial down-
sampling factor by f. Reconstruction FIDs [29] and PSNR are
calculated on ImageNet-val. [12]; see also Tab. 8.
results in image synthesis [30,85] and beyond [7,45,48,57],
and define the state-of-the-art in class-conditional image
synthesis [15,31] and super-resolution [72]. Moreover, even
unconditional DMs can readily be applied to tasks such
as inpainting and colorization [85] or stroke-based syn-
thesis [53], in contrast to other types of generative mod-
els [19,46,69]. Being likelihood-based models, they do not
exhibit mode-collapse and training instabilities as GANs
and, by heavily exploiting parameter sharing, they can
model highly complex distributions of natural images with-
out involving billions of parameters as in AR models [67].
Democratizing High-Resolution Image Synthesis DMs
belong to the class of likelihood-based models, whose
mode-covering behavior makes them prone to spend ex-
cessive amounts of capacity (and thus compute resources)
on modeling imperceptible details of the data [16, 73]. Al-
though the reweighted variational objective [30] aims to ad-
dress this by undersampling the initial denoising steps, DMs
are still computationally demanding, since training and
evaluating such a model requires repeated function evalu-
ations (and gradient computations) in the high-dimensional
space of RGB images. As an example, training the most
powerful DMs often takes hundreds of GPU days ( e.g. 150 -
1000 V100 days in [15]) and repeated evaluations on a noisy
version of the input space render also inference expensive,
1arXiv:2112.10752v2  [cs.CV]  13 Apr 2022
so that producing 50k samples takes approximately 5 days
[15] on a single A100 GPU. This has two consequences for
the research community and users in general: Firstly, train-
ing such a model requires massive computational resources
only available to a small fraction of the field, and leaves a
huge carbon footprint [65, 86]. Secondly, evaluating an al-
ready trained model is also expensive in time and memory,
since the same model architecture must run sequentially for
a large number of steps ( e.g. 25 - 1000 steps in [15]).
To increase the accessibility of this powerful model class
and at the same time reduce its significant resource con-
sumption, a method is needed that reduces the computa-
tional complexity for both training and sampling. Reducing
the computational demands of DMs without impairing their
performance is, therefore, key to enhance their accessibility.
Departure to Latent Space Our approach starts with
the analysis of already trained diffusion models in pixel
space: Fig. 2 shows the rate-distortion trade-off of a trained
model. As with any likelihood-based model, learning can
be roughly divided into two stages: First is a perceptual
compression stage which removes high-frequency details
but still learns little semantic variation. In the second stage,
the actual generative model learns the semantic and concep-
tual composition of the data ( semantic compression ). We
thus aim to first find a perceptually equivalent, but compu-
tationally more suitable space , in which we will train diffu-
sion models for high-resolution image synthesis.
Following common practice [11, 23, 66, 67, 96], we sep-
arate training into two distinct phases: First, we train
an autoencoder which provides a lower-dimensional (and
thereby efficient) representational space which is perceptu-
ally equivalent to the data space. Importantly, and in con-
trast to previous work [23,66], we do not need to rely on ex-
cessive spatial compression, as we train DMs in the learned
latent space, which exhibits better scaling properties with
respect to the spatial dimensionality. The reduced complex-
ity also provides efficient image generation from the latent
space with a single network pass. We dub the resulting
model class Latent Diffusion Models (LDMs).
A notable advantage of this approach is that we need to
train the universal autoencoding stage only once and can
therefore reuse it for multiple DM trainings or to explore
possibly completely different tasks [81]. This enables effi-
cient exploration of a large number of diffusion models for
various image-to-image and text-to-image tasks. For the lat-
ter, we design an architecture that connects transformers to
the DM’s UNet backbone [71] and enables arbitrary types
of token-based conditioning mechanisms, see Sec. 3.3.
In sum, our work makes the following contributions :
(i) In contrast to purely transformer-based approaches
[23, 66], our method scales more graceful to higher dimen-
sional data and can thus (a) work on a compression level
which provides more faithful and detailed reconstructions
than previous work (see Fig. 1) and (b) can be efficiently
Figure 2. Illustrating perceptual and semantic compression: Most
bits of a digital image correspond to imperceptible details. While
DMs allow to suppress this semantically meaningless information
by minimizing the responsible loss term, gradients (during train-
ing) and the neural network backbone (training and inference) still
need to be evaluated on all pixels, leading to superfluous compu-
tations and unnecessarily expensive optimization and inference.
We propose latent diffusion models (LDMs) as an effective gener-
ative model and a separate mild compression stage that only elim-
inates imperceptible details. Data and images from [30].
applied to high-resolution synthesis of megapixel images.
(ii) We achieve competitive performance on multiple
tasks (unconditional image synthesis, inpainting, stochastic
super-resolution) and datasets while significantly lowering
computational costs. Compared to pixel-based diffusion ap-
proaches, we also significantly decrease inference costs.
(iii) We show that, in contrast to previous work [93]
which learns both an encoder/decoder architecture and a
score-based prior simultaneously, our approach does not re-
quire a delicate weighting of reconstruction and generative
abilities. This ensures extremely faithful reconstructions
and requires very little regularization of the latent space.
(iv) We find that for densely conditioned tasks such
as super-resolution, inpainting and semantic synthesis, our
model can be applied in a convolutional fashion and render
large, consistent images of 10242px.
(v) Moreover, we design a general-purpose conditioning
mechanism based on cross-attention, enabling multi-modal
training. We use it to train class-conditional, text-to-image
and layout-to-image models.
(vi) Finally, we release pretrained latent diffusion
and autoencoding models at https : / / github .
com/CompVis/latent-diffusion which might be
reusable for a various tasks besides training of DMs [81].
2. Related Work
Generative Models for Image Synthesis The high di-
mensional nature of images presents distinct challenges
to generative modeling. Generative Adversarial Networks
(GAN) [27] allow for efficient sampling of high resolution
images with good perceptual quality [3, 42], but are diffi-
2
cult to optimize [2, 28, 54] and struggle to capture the full
data distribution [55]. In contrast, likelihood-based meth-
ods emphasize good density estimation which renders op-
timization more well-behaved. Variational autoencoders
(V AE) [46] and flow-based models [18, 19] enable efficient
synthesis of high resolution images [9, 44, 92], but sam-
ple quality is not on par with GANs. While autoregressive
models (ARM) [6, 10, 94, 95] achieve strong performance
in density estimation, computationally demanding architec-
tures [97] and a sequential sampling process limit them to
low resolution images. Because pixel based representations
of images contain barely perceptible, high-frequency de-
tails [16,73], maximum-likelihood training spends a dispro-
portionate amount of capacity on modeling them, resulting
in long training times. To scale to higher resolutions, several
two-stage approaches [23,67,101,103] use ARMs to model
a compressed latent image space instead of raw pixels.
Recently, Diffusion Probabilistic Models (DM) [82],
have achieved state-of-the-art results in density estimation
[45] as well as in sample quality [15]. The generative power
of these models stems from a natural fit to the inductive bi-
ases of image-like data when their underlying neural back-
bone is implemented as a UNet [15, 30, 71, 85]. The best
synthesis quality is usually achieved when a reweighted ob-
jective [30] is used for training. In this case, the DM corre-
sponds to a lossy compressor and allow to trade image qual-
ity for compression capabilities. Evaluating and optimizing
these models in pixel space, however, has the downside of
low inference speed and very high training costs. While
the former can be partially adressed by advanced sampling
strategies [47, 75, 84] and hierarchical approaches [31, 93],
training on high-resolution image data always requires to
calculate expensive gradients. We adress both drawbacks
with our proposed LDMs , which work on a compressed la-
tent space of lower dimensionality. This renders training
computationally cheaper and speeds up inference with al-
most no reduction in synthesis quality (see Fig. 1).
Two-Stage Image Synthesis To mitigate the shortcom-
ings of individual generative approaches, a lot of research
[11, 23, 67, 70, 101, 103] has gone into combining the
strengths of different methods into more efficient and per-
formant models via a two stage approach. VQ-V AEs [67,
101] use autoregressive models to learn an expressive prior
over a discretized latent space. [66] extend this approach to
text-to-image generation by learning a joint distributation
over discretized image and text representations. More gen-
erally, [70] uses conditionally invertible networks to pro-
vide a generic transfer between latent spaces of diverse do-
mains. Different from VQ-V AEs, VQGANs [23, 103] em-
ploy a first stage with an adversarial and perceptual objec-
tive to scale autoregressive transformers to larger images.
However, the high compression rates required for feasible
ARM training, which introduces billions of trainable pa-
rameters [23, 66], limit the overall performance of such ap-proaches and less compression comes at the price of high
computational cost [23, 66]. Our work prevents such trade-
offs, as our proposed LDMs scale more gently to higher
dimensional latent spaces due to their convolutional back-
bone. Thus, we are free to choose the level of compression
which optimally mediates between learning a powerful first
stage, without leaving too much perceptual compression up
to the generative diffusion model while guaranteeing high-
fidelity reconstructions (see Fig. 1).
While approaches to jointly [93] or separately [80] learn
an encoding/decoding model together with a score-based
prior exist, the former still require a difficult weighting be-
tween reconstruction and generative capabilities [11] and
are outperformed by our approach (Sec. 4), and the latter
focus on highly structured images such as human faces.
3. Method
To lower the computational demands of training diffu-
sion models towards high-resolution image synthesis, we
observe that although diffusion models allow to ignore
perceptually irrelevant details by undersampling the corre-
sponding loss terms [30], they still require costly function
evaluations in pixel space, which causes huge demands in
computation time and energy resources.
We propose to circumvent this drawback by introducing
an explicit separation of the compressive from the genera-
tive learning phase (see Fig. 2). To achieve this, we utilize
an autoencoding model which learns a space that is percep-
tually equivalent to the image space, but offers significantly
reduced computational complexity.
Such an approach offers several advantages: (i) By leav-
ing the high-dimensional image space, we obtain DMs
which are computationally much more efficient because
sampling is performed on a low-dimensional space. (ii) We
exploit the inductive bias of DMs inherited from their UNet
architecture [71], which makes them particularly effective
for data with spatial structure and therefore alleviates the
need for aggressive, quality-reducing compression levels as
required by previous approaches [23, 66]. (iii) Finally, we
obtain general-purpose compression models whose latent
space can be used to train multiple generative models and
which can also be utilized for other downstream applica-
tions such as single-image CLIP-guided synthesis [25].
3.1. Perceptual Image Compression
Our perceptual compression model is based on previous
work [23] and consists of an autoencoder trained by com-
bination of a perceptual loss [106] and a patch-based [33]
adversarial objective [20, 23, 103]. This ensures that the re-
constructions are confined to the image manifold by enforc-
ing local realism and avoids bluriness introduced by relying
solely on pixel-space losses such as L2orL1objectives.
More precisely, given an image x2RHW3in RGB
space, the encoder Eencodesxinto a latent representa-
3
tionz=E(x), and the decoder Dreconstructs the im-
age from the latent, giving ~x=D(z) =D(E(x)), where
z2Rhwc. Importantly, the encoder downsamples the
image by a factor f=H=h =W=w , and we investigate
different downsampling factors f= 2m, withm2N.
In order to avoid arbitrarily high-variance latent spaces,
we experiment with two different kinds of regularizations.
The first variant, KL-reg. , imposes a slight KL-penalty to-
wards a standard normal on the learned latent, similar to a
V AE [46, 69], whereas VQ-reg. uses a vector quantization
layer [96] within the decoder. This model can be interpreted
as a VQGAN [23] but with the quantization layer absorbed
by the decoder. Because our subsequent DM is designed
to work with the two-dimensional structure of our learned
latent space z=E(x), we can use relatively mild compres-
sion rates and achieve very good reconstructions. This is
in contrast to previous works [23, 66], which relied on an
arbitrary 1D ordering of the learned space zto model its
distribution autoregressively and thereby ignored much of
the inherent structure of z. Hence, our compression model
preserves details of xbetter (see Tab. 8). The full objective
and training details can be found in the supplement.
3.2. Latent Diffusion Models
Diffusion Models [82] are probabilistic models designed to
learn a data distribution p(x)by gradually denoising a nor-
mally distributed variable, which corresponds to learning
the reverse process of a fixed Markov Chain of length T.
For image synthesis, the most successful models [15,30,72]
rely on a reweighted variant of the variational lower bound
onp(x), which mirrors denoising score-matching [85].
These models can be interpreted as an equally weighted
sequence of denoising autoencoders (xt;t);t= 1:::T ,
which are trained to predict a denoised variant of their input
xt, wherextis a noisy version of the input x. The corre-
sponding objective can be simplified to (Sec. B)
LDM=Ex;N(0;1);th
k(xt;t)k2
2i
; (1)
withtuniformly sampled from f1;:::;Tg.
Generative Modeling of Latent Representations With
our trained perceptual compression models consisting of E
andD, we now have access to an efficient, low-dimensional
latent space in which high-frequency, imperceptible details
are abstracted away. Compared to the high-dimensional
pixel space, this space is more suitable for likelihood-based
generative models, as they can now (i) focus on the impor-
tant, semantic bits of the data and (ii) train in a lower di-
mensional, computationally much more efficient space.
Unlike previous work that relied on autoregressive,
attention-based transformer models in a highly compressed,
discrete latent space [23,66,103], we can take advantage of
image-specific inductive biases that our model offers. This
Semantic  
 Map
crossattentionLatent SpaceConditioning  
TextDiffusion Process
denoising step switch skip connectionRepres  
entations
Pixel SpaceImagesDenoising U-Net
concatFigure 3. We condition LDMs either via concatenation or by a
more general cross-attention mechanism. See Sec. 3.3
includes the ability to build the underlying UNet primar-
ily from 2D convolutional layers, and further focusing the
objective on the perceptually most relevant bits using the
reweighted bound, which now reads
LLDM :=EE(x);N(0;1);th
k(zt;t)k2
2i
:(2)
The neural backbone (;t)of our model is realized as a
time-conditional UNet [71]. Since the forward process is
fixed,ztcan be efficiently obtained from Eduring training,
and samples from p(z) can be decoded to image space with
a single pass through D.
3.3. Conditioning Mechanisms
Similar to other types of generative models [56, 83],
diffusion models are in principle capable of modeling
conditional distributions of the form p(zjy). This can
be implemented with a conditional denoising autoencoder
(zt;t;y)and paves the way to controlling the synthesis
process through inputs ysuch as text [68], semantic maps
[33, 61] or other image-to-image translation tasks [34].
In the context of image synthesis, however, combining
the generative power of DMs with other types of condition-
ings beyond class-labels [15] or blurred variants of the input
image [72] is so far an under-explored area of research.
We turn DMs into more flexible conditional image gener-
ators by augmenting their underlying UNet backbone with
the cross-attention mechanism [97], which is effective for
learning attention-based models of various input modali-
ties [35,36]. To pre-process yfrom various modalities (such
as language prompts) we introduce a domain specific en-
coderthat projects yto an intermediate representation
(y)2RMd, which is then mapped to the intermediate
layers of the UNet via a cross-attention layer implementing
Attention (Q;K;V ) =softmax
QKT
p
d

V, with
Q=W(i)
Q
'i(zt); K=W(i)
K
(y); V=W(i)
V
(y):
Here,'i(zt)2RNdi
denotes a (flattened) intermediate
representation of the UNet implementing andW(i)
V2
4
CelebAHQ FFHQ LSUN-Churches LSUN-Beds ImageNet
Figure 4. Samples from LDMs trained on CelebAHQ [39], FFHQ [41], LSUN-Churches [102], LSUN-Bedrooms [102] and class-
conditional ImageNet [12], each with a resolution of 256256. Best viewed when zoomed in. For more samples cf. the supplement.
Rddi
,W(i)
Q2Rdd&W(i)
K2Rddare learnable pro-
jection matrices [36, 97]. See Fig. 3 for a visual depiction.
Based on image-conditioning pairs, we then learn the
conditional LDM via
LLDM :=EE(x);y;N(0;1);th
k(zt;t;(y))k2
2i
;(3)
where bothandare jointly optimized via Eq. 3. This
conditioning mechanism is flexible as can be parameter-
ized with domain-specific experts, e.g. (unmasked) trans-
formers [97] when yare text prompts (see Sec. 4.3.1)
4. Experiments
LDMs provide means to flexible and computationally
tractable diffusion based image synthesis of various image
modalities, which we empirically show in the following.
Firstly, however, we analyze the gains of our models com-
pared to pixel-based diffusion models in both training and
inference. Interestingly, we find that LDMs trained in VQ-
regularized latent spaces sometimes achieve better sample
quality, even though the reconstruction capabilities of VQ-
regularized first stage models slightly fall behind those of
their continuous counterparts, cf. Tab. 8. A visual compari-
son between the effects of first stage regularization schemes
onLDM training and their generalization abilities to resolu-
tions>2562can be found in Appendix D.1. In E.2 we list
details on architecture, implementation, training and evalu-
ation for all results presented in this section.
4.1. On Perceptual Compression Tradeoffs
This section analyzes the behavior of our LDMs with dif-
ferent downsampling factors f2f1;2;4;8;16;32g(abbre-
viated as LDM-f, where LDM-1 corresponds to pixel-based
DMs). To obtain a comparable test-field, we fix the com-
putational resources to a single NVIDIA A100 for all ex-
periments in this section and train all models for the same
number of steps and with the same number of parameters.
Tab. 8 shows hyperparameters and reconstruction perfor-
mance of the first stage models used for the LDMs com-pared in this section. Fig. 6 shows sample quality as a func-
tion of training progress for 2M steps of class-conditional
models on the ImageNet [12] dataset. We see that, i) small
downsampling factors for LDM-f1,2gresult in slow train-
ing progress, whereas ii) overly large values of fcause stag-
nating fidelity after comparably few training steps. Revis-
iting the analysis above (Fig. 1 and 2) we attribute this to
i) leaving most of perceptual compression to the diffusion
model and ii) too strong first stage compression resulting
in information loss and thus limiting the achievable qual-
ity.LDM-f4-16gstrike a good balance between efficiency
and perceptually faithful results, which manifests in a sig-
nificant FID [29] gap of 38 between pixel-based diffusion
(LDM-1 ) and LDM-8 after 2M training steps.
In Fig. 7, we compare models trained on CelebA-
HQ [39] and ImageNet in terms sampling speed for differ-
ent numbers of denoising steps with the DDIM sampler [84]
and plot it against FID-scores [29]. LDM-f4-8goutper-
form models with unsuitable ratios of perceptual and con-
ceptual compression. Especially compared to pixel-based
LDM-1 , they achieve much lower FID scores while simulta-
neously significantly increasing sample throughput. Com-
plex datasets such as ImageNet require reduced compres-
sion rates to avoid reducing quality. In summary, LDM-4
and-8offer the best conditions for achieving high-quality
synthesis results.
4.2. Image Generation with Latent Diffusion
We train unconditional models of 2562images on
CelebA-HQ [39], FFHQ [41], LSUN-Churches and
-Bedrooms [102] and evaluate the i) sample quality and ii)
their coverage of the data manifold using ii) FID [29] and
ii) Precision-and-Recall [50]. Tab. 1 summarizes our re-
sults. On CelebA-HQ, we report a new state-of-the-art FID
of5:11, outperforming previous likelihood-based models as
well as GANs. We also outperform LSGM [93] where a la-
tent diffusion model is trained jointly together with the first
stage. In contrast, we train diffusion models in a fixed space
5
Text-to-Image Synthesis on LAION. 1.45B Model.
’A street sign that reads
“Latent Diffusion” ’’A zombie in the
style of Picasso’’An image of an animal
half mouse half octopus’’An illustration of a slightly
conscious neural network’’A painting of a
squirrel eating a burger’’A watercolor painting of a
chair that looks like an octopus’’A shirt with the inscription:
“I love generative models!” ’
Figure 5. Samples for user-defined text prompts from our model for text-to-image synthesis, LDM-8 (KL) , which was trained on the
LAION [78] database. Samples generated with 200 DDIM steps and = 1:0. We use unconditional guidance [32] with s= 10:0.
Figure 6. Analyzing the training of class-conditional LDMs with
different downsampling factors fover 2M train steps on the Im-
ageNet dataset. Pixel-based LDM-1 requires substantially larger
train times compared to models with larger downsampling factors
(LDM-f4-16g). Too much perceptual compression as in LDM-32
limits the overall sample quality. All models are trained on a sin-
gle NVIDIA A100 with the same computational budget. Results
obtained with 100 DDIM steps [84] and = 0.
Figure 7. Comparing LDMs with varying compression on the
CelebA-HQ (left) and ImageNet (right) datasets. Different mark-
ers indicatef10;20;50;100;200gsampling steps using DDIM,
from right to left along each line. The dashed line shows the FID
scores for 200 steps, indicating the strong performance of LDM-
f4-8g. FID scores assessed on 5000 samples. All models were
trained for 500k (CelebA) / 2M (ImageNet) steps on an A100.
and avoid the difficulty of weighing reconstruction quality
against learning the prior over the latent space, see Fig. 1-2.
We outperform prior diffusion based approaches on all
but the LSUN-Bedrooms dataset, where our score is close
to ADM [15], despite utilizing half its parameters and re-
quiring 4-times less train resources (see Appendix E.3.5).CelebA-HQ 256256 FFHQ 256256
Method FID# Prec." Recall" Method FID# Prec." Recall"
DC-V AE [63] 15.8 - - ImageBART [21] 9.57 - -
VQGAN+T. [23] (k=400) 10.2 - - U-Net GAN (+aug) [77] 10.9 (7.6) - -
PGGAN [39] 8.0 - - UDM [43] 5.54 - -
LSGM [93] 7.22 - - StyleGAN [41] 4.16 0.71 0.46
UDM [43] 7.16 - - ProjectedGAN [76] 3.08 0.65 0.46
LDM-4 (ours, 500-sy) 5.11 0.72 0.49 LDM-4 (ours, 200-s) 4.98 0.73 0.50
LSUN-Churches 256256 LSUN-Bedrooms 256256
Method FID# Prec." Recall" Method FID# Prec." Recall"
DDPM [30] 7.89 - - ImageBART [21] 5.51 - -
ImageBART [21] 7.32 - - DDPM [30] 4.9 - -
PGGAN [39] 6.42 - - UDM [43] 4.57 - -
StyleGAN [41] 4.21 - - StyleGAN [41] 2.35 0.59 0.48
StyleGAN2 [42] 3.86 - - ADM [15] 1.90 0.66 0.51
ProjectedGAN [76] 1.59 0.61 0.44 ProjectedGAN [76] 1.52 0.61 0.34
LDM-8(ours, 200-s) 4.02 0.64 0.52 LDM-4 (ours, 200-s) 2.95 0.66 0.48
Table 1. Evaluation metrics for unconditional image synthesis.
CelebA-HQ results reproduced from [43, 63, 100], FFHQ from
[42, 43].y:N-s refers toNsampling steps with the DDIM [84]
sampler.: trained in KL-regularized latent space. Additional re-
sults can be found in the supplementary.
Text-Conditional Image Synthesis
Method FID# IS" Nparams
CogViewy[17] 27.10 18.20 4B self-ranking, rejection rate 0.017
LAFITEy[109] 26.94 26.02 75M
GLIDE[59] 12.24 - 6B 277 DDIM steps, c.f.g. [32] s= 3
Make-A-Scene[26] 11.84 - 4B c.f.g for AR models [98] s= 5
LDM-KL-8 23.31 20.03 0.33 1.45B 250 DDIM steps
LDM-KL-8-G12.63 30.29 0.42 1.45B 250 DDIM steps, c.f.g. [32] s= 1:5
Table 2. Evaluation of text-conditional image synthesis on the
256256-sized MS-COCO [51] dataset: with 250 DDIM [84]
steps our model is on par with the most recent diffusion [59] and
autoregressive [26] methods despite using significantly less pa-
rameters.y/:Numbers from [109]/ [26]
Moreover, LDMs consistently improve upon GAN-based
methods in Precision and Recall, thus confirming the ad-
vantages of their mode-covering likelihood-based training
objective over adversarial approaches. In Fig. 4 we also
show qualitative results on each dataset.
6
Figure 8. Layout-to-image synthesis with an LDM on COCO [4],
see Sec. 4.3.1. Quantitative evaluation in the supplement D.3.
4.3. Conditional Latent Diffusion
4.3.1 Transformer Encoders for LDMs
By introducing cross-attention based conditioning into
LDMs we open them up for various conditioning modali-
ties previously unexplored for diffusion models. For text-
to-image image modeling, we train a 1.45B parameter
KL-regularized LDM conditioned on language prompts on
LAION-400M [78]. We employ the BERT-tokenizer [14]
and implement as a transformer [97] to infer a latent
code which is mapped into the UNet via (multi-head) cross-
attention (Sec. 3.3). This combination of domain specific
experts for learning a language representation and visual
synthesis results in a powerful model, which generalizes
well to complex, user-defined text prompts, cf. Fig. 8 and 5.
For quantitative analysis, we follow prior work and evaluate
text-to-image generation on the MS-COCO [51] validation
set, where our model improves upon powerful AR [17, 66]
and GAN-based [109] methods, cf. Tab. 2. We note that ap-
plying classifier-free diffusion guidance [32] greatly boosts
sample quality, such that the guided LDM-KL-8-G is on par
with the recent state-of-the-art AR [26] and diffusion mod-
els [59] for text-to-image synthesis, while substantially re-
ducing parameter count. To further analyze the flexibility of
the cross-attention based conditioning mechanism we also
train models to synthesize images based on semantic lay-
outs on OpenImages [49], and finetune on COCO [4], see
Fig. 8. See Sec. D.3 for the quantitative evaluation and im-
plementation details.
Lastly, following prior work [3, 15, 21, 23], we evalu-
ate our best-performing class-conditional ImageNet mod-
els withf2 f4;8gfrom Sec. 4.1 in Tab. 3, Fig. 4 and
Sec. D.4. Here we outperform the state of the art diffu-
sion model ADM [15] while significantly reducing compu-
tational requirements and parameter count, cf. Tab 18.
4.3.2 Convolutional Sampling Beyond 2562
By concatenating spatially aligned conditioning informa-
tion to the input of ,LDMs can serve as efficient general-Method FID# IS" Precision" Recall"Nparams
BigGan-deep [3] 6.95 203.6 2.6 0.87 0.28 340M -
ADM [15] 10.94 100.98 0.69 0.63 554M 250 DDIM steps
ADM-G [15] 4.59 186.7 0.82 0.52 608M 250 DDIM steps
LDM-4 (ours) 10.56 103.49 1.24 0.71 0.62 400M 250 DDIM steps
LDM-4 -G (ours) 3.60 247.67 5.59 0.87 0.48 400M 250 steps, c.f.g [32], s= 1:5
Table 3. Comparison of a class-conditional ImageNet LDM with
recent state-of-the-art methods for class-conditional image gener-
ation on ImageNet [12]. A more detailed comparison with addi-
tional baselines can be found in D.4, Tab. 10 and F. c.f.g. denotes
classifier-free guidance with a scale sas proposed in [32].
purpose image-to-image translation models. We use this
to train models for semantic synthesis, super-resolution
(Sec. 4.4) and inpainting (Sec. 4.5). For semantic synthe-
sis, we use images of landscapes paired with semantic maps
[23, 61] and concatenate downsampled versions of the se-
mantic maps with the latent image representation of a f= 4
model (VQ-reg., see Tab. 8). We train on an input resolution
of2562(crops from 3842) but find that our model general-
izes to larger resolutions and can generate images up to the
megapixel regime when evaluated in a convolutional man-
ner (see Fig. 9). We exploit this behavior to also apply the
super-resolution models in Sec. 4.4 and the inpainting mod-
els in Sec. 4.5 to generate large images between 5122and
10242. For this application, the signal-to-noise ratio (in-
duced by the scale of the latent space) significantly affects
the results. In Sec. D.1 we illustrate this when learning an
LDM on (i) the latent space as provided by a f= 4 model
(KL-reg., see Tab. 8), and (ii) a rescaled version, scaled by
the component-wise standard deviation.
The latter, in combination with classifier-free guid-
ance [32], also enables the direct synthesis of >2562im-
ages for the text-conditional LDM-KL-8-G as in Fig. 13.
Figure 9. A LDM trained on 2562resolution can generalize to
larger resolution (here: 5121024 ) for spatially conditioned tasks
such as semantic synthesis of landscape images. See Sec. 4.3.2.
4.4. Super-Resolution with Latent Diffusion
LDMs can be efficiently trained for super-resolution by
diretly conditioning on low-resolution images via concate-
nation ( cf. Sec. 3.3). In a first experiment, we follow SR3
7
bicubic LDM -SR SR3
Figure 10. ImageNet 64 !256 super-resolution on ImageNet-Val.
LDM-SR has advantages at rendering realistic textures but SR3
can synthesize more coherent fine structures. See appendix for
additional samples and cropouts. SR3 results from [72].
[72] and fix the image degradation to a bicubic interpola-
tion with 4-downsampling and train on ImageNet follow-
ing SR3’s data processing pipeline. We use the f= 4 au-
toencoding model pretrained on OpenImages (VQ-reg., cf.
Tab. 8) and concatenate the low-resolution conditioning y
and the inputs to the UNet, i.e.is the identity. Our quali-
tative and quantitative results (see Fig. 10 and Tab. 5) show
competitive performance and LDM-SR outperforms SR3
in FID while SR3 has a better IS. A simple image regres-
sion model achieves the highest PSNR and SSIM scores;
however these metrics do not align well with human per-
ception [106] and favor blurriness over imperfectly aligned
high frequency details [72]. Further, we conduct a user
study comparing the pixel-baseline with LDM-SR. We fol-
low SR3 [72] where human subjects were shown a low-res
image in between two high-res images and asked for pref-
erence. The results in Tab. 4 affirm the good performance
of LDM-SR. PSNR and SSIM can be pushed by using a
post-hoc guiding mechanism [15] and we implement this
image-based guider via a perceptual loss, see Sec. D.6.
SR on ImageNet Inpainting on Places
User Study Pixel-DM ( f1)LDM-4 LAMA [88] LDM-4
Task 1: Preference vs GT" 16.0% 30.4% 13.6% 21.0%
Task 2: Preference Score" 29.4% 70.6% 31.9% 68.1%
Table 4. Task 1: Subjects were shown ground truth and generated
image and asked for preference. Task 2: Subjects had to decide
between two generated images. More details in E.3.6
Since the bicubic degradation process does not generalize
well to images which do not follow this pre-processing, we
also train a generic model, LDM-BSR , by using more di-
verse degradation. The results are shown in Sec. D.6.1.Method FID# IS" PSNR" SSIM"Nparams [samples
s]()
Image Regression [72] 15.2 121.1 27.9 0.801 625M N/A
SR3 [72] 5.2 180.1 26.4 0.762 625M N/A
LDM-4 (ours, 100 steps) 2.8y/4.8z166.3 24.4 3.8 0.690.14 169M 4.62
emphLDM-4 (ours, big, 100 steps) 2.4y/4.3z174.9 24.74.1 0.710.15 552M 4.5
LDM-4 (ours, 50 steps, guiding) 4.4y/6.4z153.7 25.8 3.7 0.740.12 184M 0.38
Table 5.4upscaling results on ImageNet-Val. ( 2562);y: FID
features computed on validation split,z: FID features computed
on train split;: Assessed on a NVIDIA A100
train throughput sampling throughputytrain+val FID@2k
Model (reg.-type) samples/sec. @256 @512 hours/epoch epoch 6
LDM-1 (no first stage) 0.11 0.26 0.07 20.66 24.74
LDM-4 (KL, w/ attn) 0.32 0.97 0.34 7.66 15.21
LDM-4 (VQ, w/ attn) 0.33 0.97 0.34 7.04 14.99
LDM-4 (VQ, w/o attn) 0.35 0.99 0.36 6.66 15.95
Table 6. Assessing inpainting efficiency.y: Deviations from Fig. 7
due to varying GPU settings/batch sizes cf. the supplement.
4.5. Inpainting with Latent Diffusion
Inpainting is the task of filling masked regions of an im-
age with new content either because parts of the image are
are corrupted or to replace existing but undesired content
within the image. We evaluate how our general approach
for conditional image generation compares to more special-
ized, state-of-the-art approaches for this task. Our evalua-
tion follows the protocol of LaMa [88], a recent inpainting
model that introduces a specialized architecture relying on
Fast Fourier Convolutions [8]. The exact training & evalua-
tion protocol on Places [108] is described in Sec. E.2.2.
We first analyze the effect of different design choices for
the first stage. In particular, we compare the inpainting ef-
ficiency of LDM-1 (i.e. a pixel-based conditional DM) with
LDM-4 , for both KLandVQregularizations, as well as VQ-
LDM-4 without any attention in the first stage (see Tab. 8),
where the latter reduces GPU memory for decoding at high
resolutions. For comparability, we fix the number of param-
eters for all models. Tab. 6 reports the training and sampling
throughput at resolution 2562and5122, the total training
time in hours per epoch and the FID score on the validation
split after six epochs. Overall, we observe a speed-up of at
least2:7between pixel- and latent-based diffusion models
while improving FID scores by a factor of at least 1:6.
The comparison with other inpainting approaches in
Tab. 7 shows that our model with attention improves the
overall image quality as measured by FID over that of [88].
LPIPS between the unmasked images and our samples is
slightly higher than that of [88]. We attribute this to [88]
only producing a single result which tends to recover more
of an average image compared to the diverse results pro-
duced by our LDM cf. Fig. 21. Additionally in a user study
(Tab. 4) human subjects favor our results over those of [88].
Based on these initial results, we also trained a larger dif-
fusion model ( bigin Tab. 7) in the latent space of the VQ-
regularized first stage without attention. Following [15],
the UNet of this diffusion model uses attention layers on
three levels of its feature hierarchy, the BigGAN [3] residual
block for up- and downsampling and has 387M parameters
8
input result
Figure 11. Qualitative results on object removal with our big, w/
ftinpainting model. For more results, see Fig. 22.
instead of 215M. After training, we noticed a discrepancy
in the quality of samples produced at resolutions 2562and
5122, which we hypothesize to be caused by the additional
attention modules. However, fine-tuning the model for half
an epoch at resolution 5122allows the model to adjust to
the new feature statistics and sets a new state of the art FID
on image inpainting ( big, w/o attn, w/ ft in Tab. 7, Fig. 11.).
5. Limitations & Societal Impact
Limitations While LDMs significantly reduce computa-
tional requirements compared to pixel-based approaches,
their sequential sampling process is still slower than that
of GANs. Moreover, the use of LDMs can be question-
able when high precision is required: although the loss of
image quality is very small in our f= 4autoencoding mod-
els (see Fig. 1), their reconstruction capability can become
a bottleneck for tasks that require fine-grained accuracy in
pixel space. We assume that our superresolution models
(Sec. 4.4) are already somewhat limited in this respect.
Societal Impact Generative models for media like im-
agery are a double-edged sword: On the one hand, they40-50% masked All samples
Method FID# LPIPS# FID# LPIPS#
LDM-4 (ours, big, w/ ft) 9.39 0.246 0.042 1.50 0.137 0.080
LDM-4 (ours, big, w/o ft) 12.89 0.257 0.047 2.40 0.142 0.085
LDM-4 (ours, w/ attn) 11.87 0.257 0.042 2.15 0.144 0.084
LDM-4 (ours, w/o attn) 12.60 0.259 0.041 2.37 0.145 0.084
LaMa [88]y12.31 0.243 0.038 2.23 0.134 0.080
LaMa [88] 12.0 0.24 2.21 0.14
CoModGAN [107] 10.4 0.26 1.82 0.15
RegionWise [52] 21.3 0.27 4.75 0.15
DeepFill v2 [104] 22.1 0.28 5.20 0.16
EdgeConnect [58] 30.5 0.28 8.37 0.16
Table 7. Comparison of inpainting performance on 30k crops of
size512512from test images of Places [108]. The column 40-
50% reports metrics computed over hard examples where 40-50%
of the image region have to be inpainted.yrecomputed on our test
set, since the original test set used in [88] was not available.
enable various creative applications, and in particular ap-
proaches like ours that reduce the cost of training and in-
ference have the potential to facilitate access to this tech-
nology and democratize its exploration. On the other hand,
it also means that it becomes easier to create and dissemi-
nate manipulated data or spread misinformation and spam.
In particular, the deliberate manipulation of images (“deep
fakes”) is a common problem in this context, and women in
particular are disproportionately affected by it [13, 24].
Generative models can also reveal their training data
[5, 90], which is of great concern when the data contain
sensitive or personal information and were collected with-
out explicit consent. However, the extent to which this also
applies to DMs of images is not yet fully understood.
Finally, deep learning modules tend to reproduce or ex-
acerbate biases that are already present in the data [22, 38,
91]. While diffusion models achieve better coverage of the
data distribution than e.g. GAN-based approaches, the ex-
tent to which our two-stage approach that combines adver-
sarial training and a likelihood-based objective misrepre-
sents the data remains an important research question.
For a more general, detailed discussion of the ethical
considerations of deep generative models, see e.g. [13].
6. Conclusion
We have presented latent diffusion models, a simple and
efficient way to significantly improve both the training and
sampling efficiency of denoising diffusion models with-
out degrading their quality. Based on this and our cross-
attention conditioning mechanism, our experiments could
demonstrate favorable results compared to state-of-the-art
methods across a wide range of conditional image synthesis
tasks without task-specific architectures.
This work has been supported by the German Federal Ministry for
Economic Affairs and Energy within the project ’KI-Absicherung - Safe
AI for automated driving’ and by the German Research Foundation (DFG)
project 421703927.
9
References
[1] Eirikur Agustsson and Radu Timofte. NTIRE 2017 chal-
lenge on single image super-resolution: Dataset and study.
In2017 IEEE Conference on Computer Vision and Pattern
Recognition Workshops, CVPR Workshops 2017, Honolulu,
HI, USA, July 21-26, 2017 , pages 1122–1131. IEEE Com-
puter Society, 2017. 1
[2] Martin Arjovsky, Soumith Chintala, and L ´eon Bottou.
Wasserstein gan, 2017. 3
[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large
scale GAN training for high fidelity natural image synthe-
sis. In Int. Conf. Learn. Represent. , 2019. 1, 2, 7, 8, 22,
28
[4] Holger Caesar, Jasper R. R. Uijlings, and Vittorio Ferrari.
Coco-stuff: Thing and stuff classes in context. In 2018
IEEE Conference on Computer Vision and Pattern Recog-
nition, CVPR 2018, Salt Lake City, UT, USA, June 18-
22, 2018 , pages 1209–1218. Computer Vision Foundation /
IEEE Computer Society, 2018. 7, 20, 22
[5] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew
Jagielski, Ariel Herbert-V oss, Katherine Lee, Adam
Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al.
Extracting training data from large language models. In
30th USENIX Security Symposium (USENIX Security 21) ,
pages 2633–2650, 2021. 9
[6] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Hee-
woo Jun, David Luan, and Ilya Sutskever. Generative pre-
training from pixels. In ICML , volume 119 of Proceedings
of Machine Learning Research , pages 1691–1703. PMLR,
2020. 3
[7] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss, Mo-
hammad Norouzi, and William Chan. Wavegrad: Estimat-
ing gradients for waveform generation. In ICLR . OpenRe-
view.net, 2021. 1
[8] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolu-
tion. In NeurIPS , 2020. 8
[9] Rewon Child. Very deep vaes generalize autoregressive
models and can outperform them on images. CoRR ,
abs/2011.10650, 2020. 3
[10] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever.
Generating long sequences with sparse transformers.
CoRR , abs/1904.10509, 2019. 3
[11] Bin Dai and David P. Wipf. Diagnosing and enhancing V AE
models. In ICLR (Poster) . OpenReview.net, 2019. 2, 3
[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li,
and Fei-Fei Li. Imagenet: A large-scale hierarchical im-
age database. In CVPR , pages 248–255. IEEE Computer
Society, 2009. 1, 5, 7, 22
[13] Emily Denton. Ethical considerations of generative ai. AI
for Content Creation Workshop, CVPR, 2021. 9
[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. BERT: pre-training of deep bidirec-
tional transformers for language understanding. CoRR ,
abs/1810.04805, 2018. 7
[15] Prafulla Dhariwal and Alex Nichol. Diffusion models beat
gans on image synthesis. CoRR , abs/2105.05233, 2021. 1,
2, 3, 4, 6, 7, 8, 18, 22, 25, 26, 28[16] Sander Dieleman. Musings on typicality, 2020. 1, 3
[17] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng,
Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao,
Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-
image generation via transformers. CoRR , abs/2105.13290,
2021. 6, 7
[18] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice:
Non-linear independent components estimation, 2015. 3
[19] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Ben-
gio. Density estimation using real NVP. In 5th Inter-
national Conference on Learning Representations, ICLR
2017, Toulon, France, April 24-26, 2017, Conference Track
Proceedings . OpenReview.net, 2017. 1, 3
[20] Alexey Dosovitskiy and Thomas Brox. Generating images
with perceptual similarity metrics based on deep networks.
In Daniel D. Lee, Masashi Sugiyama, Ulrike von Luxburg,
Isabelle Guyon, and Roman Garnett, editors, Adv. Neural
Inform. Process. Syst. , pages 658–666, 2016. 3
[21] Patrick Esser, Robin Rombach, Andreas Blattmann, and
Bj¨orn Ommer. Imagebart: Bidirectional context with multi-
nomial diffusion for autoregressive image synthesis. CoRR ,
abs/2108.08827, 2021. 6, 7, 22
[22] Patrick Esser, Robin Rombach, and Bj ¨orn Ommer. A
note on data biases in generative models. arXiv preprint
arXiv:2012.02516 , 2020. 9
[23] Patrick Esser, Robin Rombach, and Bj ¨orn Ommer. Taming
transformers for high-resolution image synthesis. CoRR ,
abs/2012.09841, 2020. 2, 3, 4, 6, 7, 21, 22, 29, 34, 36
[24] Mary Anne Franks and Ari Ezra Waldman. Sex, lies, and
videotape: Deep fakes and free speech delusions. Md. L.
Rev., 78:892, 2018. 9
[25] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. Clipdraw:
Exploring text-to-drawing synthesis through language-
image encoders. ArXiv , abs/2106.14843, 2021. 3
[26] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin,
Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-
based text-to-image generation with human priors. CoRR ,
abs/2203.13131, 2022. 6, 7, 16
[27] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing
Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville,
and Yoshua Bengio. Generative adversarial networks.
CoRR , 2014. 1, 2
[28] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent
Dumoulin, and Aaron Courville. Improved training of
wasserstein gans, 2017. 3
[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner,
Bernhard Nessler, and Sepp Hochreiter. Gans trained by
a two time-scale update rule converge to a local nash equi-
librium. In Adv. Neural Inform. Process. Syst. , pages 6626–
6637, 2017. 1, 5, 26
[30] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif-
fusion probabilistic models. In NeurIPS , 2020. 1, 2, 3, 4,
6, 17
[31] Jonathan Ho, Chitwan Saharia, William Chan, David J.
Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded
diffusion models for high fidelity image generation. CoRR ,
abs/2106.15282, 2021. 1, 3, 22
10
[32] Jonathan Ho and Tim Salimans. Classifier-free diffusion
guidance. In NeurIPS 2021 Workshop on Deep Generative
Models and Downstream Applications , 2021. 6, 7, 16, 22,
28, 37, 38
[33] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A.
Efros. Image-to-image translation with conditional adver-
sarial networks. In CVPR , pages 5967–5976. IEEE Com-
puter Society, 2017. 3, 4
[34] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A.
Efros. Image-to-image translation with conditional adver-
sarial networks. 2017 IEEE Conference on Computer Vi-
sion and Pattern Recognition (CVPR) , pages 5967–5976,
2017. 4
[35] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste
Alayrac, Carl Doersch, Catalin Ionescu, David Ding,
Skanda Koppula, Daniel Zoran, Andrew Brock, Evan
Shelhamer, Olivier J. H ´enaff, Matthew M. Botvinick,
Andrew Zisserman, Oriol Vinyals, and Jo ˜ao Carreira.
Perceiver IO: A general architecture for structured inputs
&outputs. CoRR , abs/2107.14795, 2021. 4
[36] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals,
Andrew Zisserman, and Jo ˜ao Carreira. Perceiver: General
perception with iterative attention. In Marina Meila and
Tong Zhang, editors, Proceedings of the 38th International
Conference on Machine Learning, ICML 2021, 18-24 July
2021, Virtual Event , volume 139 of Proceedings of Machine
Learning Research , pages 4651–4664. PMLR, 2021. 4, 5
[37] Manuel Jahn, Robin Rombach, and Bj ¨orn Ommer. High-
resolution complex scene synthesis with transformers.
CoRR , abs/2105.06458, 2021. 20, 22, 27
[38] Niharika Jain, Alberto Olmo, Sailik Sengupta, Lydia
Manikonda, and Subbarao Kambhampati. Imperfect ima-
ganation: Implications of gans exacerbating biases on fa-
cial data augmentation and snapchat selfie lenses. arXiv
preprint arXiv:2001.09528 , 2020. 9
[39] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehti-
nen. Progressive growing of gans for improved quality, sta-
bility, and variation. CoRR , abs/1710.10196, 2017. 5, 6
[40] Tero Karras, Samuli Laine, and Timo Aila. A style-based
generator architecture for generative adversarial networks.
InIEEE Conf. Comput. Vis. Pattern Recog. , pages 4401–
4410, 2019. 1
[41] T. Karras, S. Laine, and T. Aila. A style-based gener-
ator architecture for generative adversarial networks. In
2019 IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition (CVPR) , 2019. 5, 6
[42] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten,
Jaakko Lehtinen, and Timo Aila. Analyzing and improv-
ing the image quality of stylegan. CoRR , abs/1912.04958,
2019. 2, 6, 28
[43] Dongjun Kim, Seungjae Shin, Kyungwoo Song, Wanmo
Kang, and Il-Chul Moon. Score matching model for un-
bounded data score. CoRR , abs/2106.05527, 2021. 6
[44] Durk P Kingma and Prafulla Dhariwal. Glow: Generative
flow with invertible 1x1 convolutions. In S. Bengio, H. Wal-
lach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R.
Garnett, editors, Advances in Neural Information Process-
ing Systems , 2018. 3[45] Diederik P. Kingma, Tim Salimans, Ben Poole, and
Jonathan Ho. Variational diffusion models. CoRR ,
abs/2107.00630, 2021. 1, 3, 16
[46] Diederik P. Kingma and Max Welling. Auto-Encoding Vari-
ational Bayes. In 2nd International Conference on Learn-
ing Representations, ICLR , 2014. 1, 3, 4, 29
[47] Zhifeng Kong and Wei Ping. On fast sampling of diffusion
probabilistic models. CoRR , abs/2106.00132, 2021. 3
[48] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and
Bryan Catanzaro. Diffwave: A versatile diffusion model
for audio synthesis. In ICLR . OpenReview.net, 2021. 1
[49] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper R. R.
Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali,
Stefan Popov, Matteo Malloci, Tom Duerig, and Vittorio
Ferrari. The open images dataset V4: unified image classi-
fication, object detection, and visual relationship detection
at scale. CoRR , abs/1811.00982, 2018. 7, 20, 22
[50] Tuomas Kynk ¨a¨anniemi, Tero Karras, Samuli Laine, Jaakko
Lehtinen, and Timo Aila. Improved precision and re-
call metric for assessing generative models. CoRR ,
abs/1904.06991, 2019. 5, 26
[51] Tsung-Yi Lin, Michael Maire, Serge J. Belongie,
Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro
Perona, Deva Ramanan, Piotr Doll ´ar, and C. Lawrence Zit-
nick. Microsoft COCO: common objects in context. CoRR ,
abs/1405.0312, 2014. 6, 7, 27
[52] Yuqing Ma, Xianglong Liu, Shihao Bai, Le-Yi Wang, Ais-
han Liu, Dacheng Tao, and Edwin Hancock. Region-wise
generative adversarial imageinpainting for large missing ar-
eas. ArXiv , abs/1909.12507, 2019. 9
[53] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, Jun-
Yan Zhu, and Stefano Ermon. Sdedit: Image synthesis
and editing with stochastic differential equations. CoRR ,
abs/2108.01073, 2021. 1
[54] Lars M. Mescheder. On the convergence properties of GAN
training. CoRR , abs/1801.04406, 2018. 3
[55] Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-
Dickstein. Unrolled generative adversarial networks. In
5th International Conference on Learning Representations,
ICLR 2017, Toulon, France, April 24-26, 2017, Conference
Track Proceedings . OpenReview.net, 2017. 3
[56] Mehdi Mirza and Simon Osindero. Conditional generative
adversarial nets. CoRR , abs/1411.1784, 2014. 4
[57] Gautam Mittal, Jesse H. Engel, Curtis Hawthorne, and Ian
Simon. Symbolic music generation with diffusion models.
CoRR , abs/2103.16091, 2021. 1
[58] Kamyar Nazeri, Eric Ng, Tony Joseph, Faisal Z. Qureshi,
and Mehran Ebrahimi. Edgeconnect: Generative im-
age inpainting with adversarial edge learning. ArXiv ,
abs/1901.00212, 2019. 9
[59] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav
Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and
Mark Chen. GLIDE: towards photorealistic image genera-
tion and editing with text-guided diffusion models. CoRR ,
abs/2112.10741, 2021. 6, 7, 16
[60] Anton Obukhov, Maximilian Seitzer, Po-Wei Wu, Se-
men Zhydenko, Jonathan Kyl, and Elvis Yu-Jing Lin.
11
High-fidelity performance metrics for generative models
in pytorch, 2020. Version: 0.3.0, DOI: 10.5281/zen-
odo.4957738. 26, 27
[61] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-
Yan Zhu. Semantic image synthesis with spatially-adaptive
normalization. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition , 2019. 4, 7
[62] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-
Yan Zhu. Semantic image synthesis with spatially-adaptive
normalization. In Proceedings of the IEEE/CVF Confer-
ence on Computer Vision and Pattern Recognition (CVPR) ,
June 2019. 22
[63] Gaurav Parmar, Dacheng Li, Kwonjoon Lee, and Zhuowen
Tu. Dual contradistinctive generative autoencoder. In IEEE
Conference on Computer Vision and Pattern Recognition,
CVPR 2021, virtual, June 19-25, 2021 , pages 823–832.
Computer Vision Foundation / IEEE, 2021. 6
[64] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On
buggy resizing libraries and surprising subtleties in fid cal-
culation. arXiv preprint arXiv:2104.11222 , 2021. 26
[65] David A. Patterson, Joseph Gonzalez, Quoc V . Le, Chen
Liang, Lluis-Miquel Munguia, Daniel Rothchild, David R.
So, Maud Texier, and Jeff Dean. Carbon emissions and
large neural network training. CoRR , abs/2104.10350,
2021. 2
[66] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott
Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya
Sutskever. Zero-shot text-to-image generation. CoRR ,
abs/2102.12092, 2021. 1, 2, 3, 4, 7, 21, 27
[67] Ali Razavi, A ¨aron van den Oord, and Oriol Vinyals. Gen-
erating diverse high-fidelity images with VQ-V AE-2. In
NeurIPS , pages 14837–14847, 2019. 1, 2, 3, 22
[68] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Lo-
geswaran, Bernt Schiele, and Honglak Lee. Generative ad-
versarial text to image synthesis. In ICML , 2016. 4
[69] Danilo Jimenez Rezende, Shakir Mohamed, and Daan
Wierstra. Stochastic backpropagation and approximate in-
ference in deep generative models. In Proceedings of the
31st International Conference on International Conference
on Machine Learning, ICML , 2014. 1, 4, 29
[70] Robin Rombach, Patrick Esser, and Bj ¨orn Ommer.
Network-to-network translation with conditional invertible
neural networks. In NeurIPS , 2020. 3
[71] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-
net: Convolutional networks for biomedical image segmen-
tation. In MICCAI (3) , volume 9351 of Lecture Notes in
Computer Science , pages 234–241. Springer, 2015. 2, 3, 4
[72] Chitwan Saharia, Jonathan Ho, William Chan, Tim Sal-
imans, David J. Fleet, and Mohammad Norouzi. Im-
age super-resolution via iterative refinement. CoRR ,
abs/2104.07636, 2021. 1, 4, 8, 16, 22, 23, 27
[73] Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P.
Kingma. Pixelcnn++: Improving the pixelcnn with dis-
cretized logistic mixture likelihood and other modifications.
CoRR , abs/1701.05517, 2017. 1, 3
[74] Dave Salvator. NVIDIA Developer Blog. https:
/ / developer . nvidia . com / blog / getting -immediate-speedups-with-a100-tf32 , 2020.
28
[75] Robin San-Roman, Eliya Nachmani, and Lior Wolf.
Noise estimation for generative diffusion models. CoRR ,
abs/2104.02600, 2021. 3
[76] Axel Sauer, Kashyap Chitta, Jens M ¨uller, and An-
dreas Geiger. Projected gans converge faster. CoRR ,
abs/2111.01007, 2021. 6
[77] Edgar Sch ¨onfeld, Bernt Schiele, and Anna Khoreva. A u-
net based discriminator for generative adversarial networks.
In2020 IEEE/CVF Conference on Computer Vision and
Pattern Recognition, CVPR 2020, Seattle, WA, USA, June
13-19, 2020 , pages 8204–8213. Computer Vision Founda-
tion / IEEE, 2020. 6
[78] Christoph Schuhmann, Richard Vencu, Romain Beaumont,
Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo
Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-
400m: Open dataset of clip-filtered 400 million image-text
pairs, 2021. 6, 7
[79] Karen Simonyan and Andrew Zisserman. Very deep con-
volutional networks for large-scale image recognition. In
Yoshua Bengio and Yann LeCun, editors, Int. Conf. Learn.
Represent. , 2015. 29, 43, 44, 45
[80] Abhishek Sinha, Jiaming Song, Chenlin Meng, and Stefano
Ermon. D2C: diffusion-denoising models for few-shot con-
ditional generation. CoRR , abs/2106.06819, 2021. 3
[81] Charlie Snell. Alien Dreams: An Emerging Art Scene.
https : / / ml . berkeley . edu / blog / posts /
clip-art/ , 2021. [Online; accessed November-2021].
2
[82] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Mah-
eswaranathan, and Surya Ganguli. Deep unsupervised
learning using nonequilibrium thermodynamics. CoRR ,
abs/1503.03585, 2015. 1, 3, 4, 18
[83] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learn-
ing structured output representation using deep conditional
generative models. In C. Cortes, N. Lawrence, D. Lee,
M. Sugiyama, and R. Garnett, editors, Advances in Neural
Information Processing Systems , volume 28. Curran Asso-
ciates, Inc., 2015. 4
[84] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois-
ing diffusion implicit models. In ICLR . OpenReview.net,
2021. 3, 5, 6, 22
[85] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma,
Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-
based generative modeling through stochastic differential
equations. CoRR , abs/2011.13456, 2020. 1, 3, 4, 18
[86] Emma Strubell, Ananya Ganesh, and Andrew McCallum.
Energy and policy considerations for modern deep learn-
ing research. In The Thirty-Fourth AAAI Conference on
Artificial Intelligence, AAAI 2020, The Thirty-Second In-
novative Applications of Artificial Intelligence Conference,
IAAI 2020, The Tenth AAAI Symposium on Educational
Advances in Artificial Intelligence, EAAI 2020, New York,
NY, USA, February 7-12, 2020 , pages 13693–13696. AAAI
Press, 2020. 2
12
[87] Wei Sun and Tianfu Wu. Learning layout and style re-
configurable gans for controllable image synthesis. CoRR ,
abs/2003.11571, 2020. 22, 27
[88] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin,
Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov,
Naejin Kong, Harshith Goka, Kiwoong Park, and Victor S.
Lempitsky. Resolution-robust large mask inpainting with
fourier convolutions. ArXiv , abs/2109.07161, 2021. 8, 9,
26, 32
[89] Tristan Sylvain, Pengchuan Zhang, Yoshua Bengio, R. De-
von Hjelm, and Shikhar Sharma. Object-centric image gen-
eration from layouts. In Thirty-Fifth AAAI Conference on
Artificial Intelligence, AAAI 2021, Thirty-Third Conference
on Innovative Applications of Artificial Intelligence, IAAI
2021, The Eleventh Symposium on Educational Advances
in Artificial Intelligence, EAAI 2021, Virtual Event, Febru-
ary 2-9, 2021 , pages 2647–2655. AAAI Press, 2021. 20,
22, 27
[90] Patrick Tinsley, Adam Czajka, and Patrick Flynn. This face
does not exist... but it might be yours! identity leakage in
generative models. In Proceedings of the IEEE/CVF Win-
ter Conference on Applications of Computer Vision , pages
1320–1328, 2021. 9
[91] Antonio Torralba and Alexei A Efros. Unbiased look at
dataset bias. In CVPR 2011 , pages 1521–1528. IEEE, 2011.
9
[92] Arash Vahdat and Jan Kautz. NV AE: A deep hierarchical
variational autoencoder. In NeurIPS , 2020. 3
[93] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-
based generative modeling in latent space. CoRR ,
abs/2106.05931, 2021. 2, 3, 5, 6
[94] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt,
koray kavukcuoglu, Oriol Vinyals, and Alex Graves. Con-
ditional image generation with pixelcnn decoders. In Ad-
vances in Neural Information Processing Systems , 2016. 3
[95] A ¨aron van den Oord, Nal Kalchbrenner, and Koray
Kavukcuoglu. Pixel recurrent neural networks. CoRR ,
abs/1601.06759, 2016. 3
[96] A ¨aron van den Oord, Oriol Vinyals, and Koray
Kavukcuoglu. Neural discrete representation learning. In
NIPS , pages 6306–6315, 2017. 2, 4, 29
[97] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser,
and Illia Polosukhin. Attention is all you need. In NIPS ,
pages 5998–6008, 2017. 3, 4, 5, 7
[98] Rivers Have Wings. Tweet on Classifier-free
guidance for autoregressive models. https :
/ / twitter . com / RiversHaveWings / status /
1478093658716966912 , 2022. 6
[99] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-
mond, Clement Delangue, Anthony Moi, Pierric Cistac,
Tim Rault, R ´emi Louf, Morgan Funtowicz, and Jamie
Brew. Huggingface’s transformers: State-of-the-art natural
language processing. CoRR , abs/1910.03771, 2019. 26
[100] Zhisheng Xiao, Karsten Kreis, Jan Kautz, and Arash Vah-
dat. V AEBM: A symbiosis between variational autoen-
coders and energy-based models. In 9th International Con-ference on Learning Representations, ICLR 2021, Virtual
Event, Austria, May 3-7, 2021 . OpenReview.net, 2021. 6
[101] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind
Srinivas. Videogpt: Video generation using VQ-V AE and
transformers. CoRR , abs/2104.10157, 2021. 3
[102] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianx-
iong Xiao. LSUN: construction of a large-scale image
dataset using deep learning with humans in the loop. CoRR ,
abs/1506.03365, 2015. 5
[103] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang,
James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge,
and Yonghui Wu. Vector-quantized image modeling with
improved vqgan, 2021. 3, 4
[104] Jiahui Yu, Zhe L. Lin, Jimei Yang, Xiaohui Shen, Xin Lu,
and Thomas S. Huang. Free-form image inpainting with
gated convolution. 2019 IEEE/CVF International Confer-
ence on Computer Vision (ICCV) , pages 4470–4479, 2019.
9
[105] K. Zhang, Jingyun Liang, Luc Van Gool, and Radu Timo-
fte. Designing a practical degradation model for deep blind
image super-resolution. ArXiv , abs/2103.14006, 2021. 23
[106] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht-
man, and Oliver Wang. The unreasonable effectiveness of
deep features as a perceptual metric. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recog-
nition (CVPR) , June 2018. 3, 8, 19
[107] Shengyu Zhao, Jianwei Cui, Yilun Sheng, Yue Dong, Xiao
Liang, Eric I-Chao Chang, and Yan Xu. Large scale image
completion via co-modulated generative adversarial net-
works. ArXiv , abs/2103.10428, 2021. 9
[108] Bolei Zhou, `Agata Lapedriza, Aditya Khosla, Aude Oliva,
and Antonio Torralba. Places: A 10 million image database
for scene recognition. IEEE Transactions on Pattern Anal-
ysis and Machine Intelligence , 40:1452–1464, 2018. 8, 9,
26
[109] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li,
Chris Tensmeyer, Tong Yu, Jiuxiang Gu, Jinhui Xu, and
Tong Sun. LAFITE: towards language-free training for
text-to-image generation. CoRR , abs/2111.13792, 2021. 6,
7, 16
13
Appendix
Figure 12. Convolutional samples from the semantic landscapes model as in Sec. 4.3.2, finetuned on 5122images.
14
’A painting of the last supper by Picasso. ’
’An oil painting of a latent space. ’’An epic painting of Gandalf the Black
summoning thunder and lightning in the mountains. ’
’A sunset over a mountain range, vector image. ’
Figure 13. Combining classifier free diffusion guidance with the convolutional sampling strategy from Sec. 4.3.2, our 1.45B parameter
text-to-image model can be used for rendering images larger than the native 2562resolution the model was trained on.
15
A. Changelog
Here we list changes between this version ( https://arxiv.org/abs/2112.10752v2 ) of the paper and the
previous version, i.e.https://arxiv.org/abs/2112.10752v1 .
• We updated the results on text-to-image synthesis in Sec. 4.3 which were obtained by training a new, larger model (1.45B
parameters). This also includes a new comparison to very recent competing methods on this task that were published on
arXiv at the same time as ( [59, 109]) or after ( [26]) the publication of our work.
• We updated results on class-conditional synthesis on ImageNet in Sec. 4.1, Tab. 3 (see also Sec. D.4) obtained by
retraining the model with a larger batch size. The corresponding qualitative results in Fig. 26 and Fig. 27 were also
updated. Both the updated text-to-image and the class-conditional model now use classifier-free guidance [32] as a
measure to increase visual fidelity.
• We conducted a user study (following the scheme suggested by Saharia et al [72]) which provides additional evaluation
for our inpainting (Sec. 4.5) and superresolution models (Sec. 4.4).
• Added Fig. 5 to the main paper, moved Fig. 18 to the appendix, added Fig. 13 to the appendix.
B. Detailed Information on Denoising Diffusion Models
Diffusion models can be specified in terms of a signal-to-noise ratio SNR (t) =2
t
2
tconsisting of sequences (t)T
t=1and
(t)T
t=1which, starting from a data sample x0, define a forward diffusion process qas
q(xtjx0) =N(xtjtx0;2
tI) (4)
with the Markov structure for s<t :
q(xtjxs) =N(xtjtjsxs;2
tjsI) (5)
tjs=t
s(6)
2
tjs=2
t2
tjs2
s (7)
Denoising diffusion models are generative models p(x0)which revert this process with a similar Markov structure running
backward in time, i.e. they are specified as
p(x0) =Z
zp(xT)TY
t=1p(xt1jxt) (8)
The evidence lower bound (ELBO) associated with this model then decomposes over the discrete time steps as
logp(x0)KL(q(xTjx0)jp(xT)) +TX
t=1Eq(xtjx0)KL(q(xt1jxt;x0)jp(xt1jxt)) (9)
The priorp(xT)is typically choosen as a standard normal distribution and the first term of the ELBO then depends only on
the final signal-to-noise ratio SNR (T). To minimize the remaining terms, a common choice to parameterize p(xt1jxt)is to
specify it in terms of the true posterior q(xt1jxt;x0)but with the unknown x0replaced by an estimate x(xt;t)based on
the current step xt. This gives [45]
p(xt1jxt):=q(xt1jxt;x(xt;t)) (10)
=N(xt1j(xt;t);2
tjt12
t1
2
tI); (11)
where the mean can be expressed as
(xt;t) =tjt12
t1
2
txt+t12
tjt1
2
tx(xt;t): (12)
16
In this case, the sum of the ELBO simplify to
TX
t=1Eq(xtjx0)KL(q(xt1jxt;x0)jp(xt1) =TX
t=1EN(j0;I)1
2(SNR(t1)SNR(t))kx0x(tx0+t;t)k2(13)
Following [30], we use the reparameterization
(xt;t) = (xttx(xt;t))=t (14)
to express the reconstruction term as a denoising objective,
kx0x(tx0+t;t)k2=2
t
2
tk(tx0+t;t)k2(15)
and the reweighting, which assigns each of the terms the same weight and results in Eq. (1).
17
C. Image Guiding Mechanisms
Samples 2562Guided Convolutional Samples 5122Convolutional Samples 5122
Figure 14. On landscapes, convolutional sampling with unconditional models can lead to homogeneous and incoherent global structures
(see column 2). L2-guiding with a low resolution image can help to reestablish coherent global structures.
An intriguing feature of diffusion models is that unconditional models can be conditioned at test-time [15, 82, 85]. In
particular, [15] presented an algorithm to guide both unconditional and conditional models trained on the ImageNet dataset
with a classifier logp(yjxt), trained on each xtof the diffusion process. We directly build on this formulation and introduce
post-hoc image-guiding :
For an epsilon-parameterized model with fixed variance, the guiding algorithm as introduced in [15] reads:
^ (zt;t) +q
12
trztlogp(yjzt): (16)
This can be interpreted as an update correcting the “score” with a conditional distribution logp(yjzt).
So far, this scenario has only been applied to single-class classification models. We re-interpret the guiding distribution
p(yjT(D(z0(zt)))) as a general purpose image-to-image translation task given a target image y, whereTcan be any
differentiable transformation adopted to the image-to-image translation task at hand, such as the identity, a downsampling
operation or similar.
18
As an example, we can assume a Gaussian guider with fixed variance 2= 1, such that
logp(yjzt) =1
2kyT(D(z0(zt)))k2
2 (17)
becomes aL2regression objective.
Fig. 14 demonstrates how this formulation can serve as an upsampling mechanism of an unconditional model trained on
2562images, where unconditional samples of size 2562guide the convolutional synthesis of 5122images and Tis a2
bicubic downsampling. Following this motivation, we also experiment with a perceptual similarity guiding and replace the
L2objective with the LPIPS [106] metric, see Sec. 4.4.
19
D. Additional Results
D.1. Choosing the Signal-to-Noise Ratio for High-Resolution Synthesis
KL-reg, w/o rescaling KL-reg, w/ rescaling VQ-reg, w/o rescaling
Figure 15. Illustrating the effect of latent space rescaling on convolutional sampling, here for semantic image synthesis on landscapes. See
Sec. 4.3.2 and Sec. D.1.
As discussed in Sec. 4.3.2, the signal-to-noise ratio induced by the variance of the latent space ( i.e. Var(z)=2
t) significantly
affects the results for convolutional sampling. For example, when training a LDM directly in the latent space of a KL-
regularized model (see Tab. 8), this ratio is very high, such that the model allocates a lot of semantic detail early on in the
reverse denoising process. In contrast, when rescaling the latent space by the component-wise standard deviation of the
latents as described in Sec. G, the SNR is descreased. We illustrate the effect on convolutional sampling for semantic image
synthesis in Fig. 15. Note that the VQ-regularized space has a variance close to 1, such that it does not have to be rescaled.
D.2. Full List of all First Stage Models
We provide a complete list of various autoenconding models trained on the OpenImages dataset in Tab. 8.
D.3. Layout-to-Image Synthesis
Here we provide the quantitative evaluation and additional samples for our layout-to-image models from Sec. 4.3.1. We
train a model on the COCO [4] and one on the OpenImages [49] dataset, which we subsequently additionally finetune on
COCO. Tab 9 shows the result. Our COCO model reaches the performance of recent state-of-the art models in layout-to-
image synthesis, when following their training and evaluation protocol [89]. When finetuning from the OpenImages model,
we surpass these works. Our OpenImages model surpasses the results of Jahn et al [37] by a margin of nearly 11 in terms of
FID. In Fig. 16 we show additional samples of the model finetuned on COCO.
D.4. Class-Conditional Image Synthesis on ImageNet
Tab. 10 contains the results for our class-conditional LDM measured in FID and Inception score (IS). LDM-8 requires
significantly fewer parameters and compute requirements (see Tab. 18) to achieve very competitive performance. Similar
to previous work, we can further boost the performance by training a classifier on each noise scale and guiding with it,
20
fjZj c R-FID# R-IS" PSNR" PSIM# SSIM"
16VQGAN [23] 16384 256 4.98 – 19.9 3:4 1.830:42 0.510:18
16VQGAN [23] 1024 256 7.94 – 19.4 3:3 1.980:43 0.500:18
8DALL-E [66] 8192 - 32.01 – 22.8 2:1 1.950:51 0.730:13
32 16384 16 31.83 40.40 1:07 17.45 2:90 2.580:48 0.410:18
16 16384 8 5.15 144.55 3:74 20.83 3:61 1.730:43 0.540:18
8 16384 4 1.14 201.92 3:97 23.07 3:99 1.170:36 0.650:16
8 256 4 1.49 194.20 3:87 22.35 3:81 1.260:37 0.620:16
4 8192 3 0.58 224.78 5:35 27.43 4:26 0.530:21 0.820:10
4y8192 3 1.06 221.94 4:58 25.21 4:17 0.720:26 0.760:12
4 256 3 0.47 223.81 4:58 26.43 4:22 0.620:24 0.800:11
2 2048 2 0.16 232.75 5:09 30.85 4:12 0.270:12 0.910:05
2 64 2 0.40 226.62 4:83 29.13 3:46 0.380:13 0.900:05
32 KL 64 2.04 189.53 3:68 22.27 3:93 1.410:40 0.610:17
32 KL 16 7.3 132.75 2:71 20.38 3:56 1.880:45 0.530:18
16 KL 16 0.87 210.31 3:97 24.08 4:22 1.070:36 0.680:15
16 KL 8 2.63 178.68 4:08 21.94 3:92 1.490:42 0.590:17
8 KL 4 0.90 209.90 4:92 24.19 4:19 1.020:35 0.690:15
4 KL 3 0.27 227.57 4:89 27.53 4:54 0.550:24 0.820:11
2 KL 2 0.086 232.66 5:16 32.47 4:19 0.200:09 0.930:04
Table 8. Complete autoencoder zoo trained on OpenImages, evaluated on ImageNet-Val. ydenotes an attention-free autoencoder.
layout-to-image synthesis on the COCO dataset
Figure 16. More samples from our best model for layout-to-image synthesis, LDM-4 , which was trained on the OpenImages dataset and
finetuned on the COCO dataset. Samples generated with 100 DDIM steps and = 0. Layouts are from the COCO validation set.
see Sec. C. Unlike the pixel-based methods, this classifier is trained very cheaply in latent space. For additional qualitative
results, see Fig. 26 and Fig. 27.
21
COCO 256256 OpenImages 256256 OpenImages 512512
Method FID# FID# FID#
LostGAN-V2 [87] 42.55 - -
OC-GAN [89] 41.65 - -
SPADE [62] 41.11 - -
VQGAN+T [37] 56.58 45.33 48.11
LDM-8 (100 steps, ours) 42.06y- -
LDM-4 (200 steps, ours) 40.9132.02 35.80
Table 9. Quantitative comparison of our layout-to-image models on the COCO [4] and OpenImages [49] datasets.y: Training from scratch
on COCO;: Finetuning from OpenImages.
Method FID# IS" Precision" Recall"Nparams
SR3 [72] 11.30 - - - 625M -
ImageBART [21] 21.19 - - - 3.5B -
ImageBART [21] 7.44 - - - 3.5B 0.05 acc. rate
VQGAN+T [23] 17.04 70.6 1.8 - - 1.3B -
VQGAN+T [23] 5.88 304.8 3.6 - - 1.3B 0.05 acc. rate
BigGan-deep [3] 6.95 203.6 2.6 0.87 0.28 340M -
ADM [15] 10.94 100.98 0.69 0.63 554M 250 DDIM steps
ADM-G [15] 4.59 186.7 0.82 0.52 608M 250 DDIM steps
ADM-G,ADM-U [15] 3.85 221.72 0.84 0.53 n/a 2 250 DDIM steps
CDM [31] 4.88 158.71 2.26 - - n/a 2 100 DDIM steps
LDM-8 (ours) 17.41 72.92 2.6 0.65 0.62 395M 200 DDIM steps, 2.9M train steps, batch size 64
LDM-8-G (ours) 8.11 190.43 2.60 0.83 0.36 506M 200 DDIM steps, classifier scale 10, 2.9M train steps, batch size 64
LDM-8 (ours) 15.51 79.03 1.03 0.65 0.63 395M 200 DDIM steps, 4.8M train steps, batch size 64
LDM-8-G (ours) 7.76 209.52 4.24 0.84 0.35 506M 200 DDIM steps, classifier scale 10, 4.8M train steps, batch size 64
LDM-4 (ours) 10.56 103.49 1.24 0.71 0.62 400M 250 DDIM steps, 178K train steps, batch size 1200
LDM-4-G (ours) 3.95 178.22 2.43 0.81 0.55 400M 250 DDIM steps, unconditional guidance [32] scale 1.25, 178K train steps, batch size 1200
LDM-4-G (ours) 3.60 247.67 5.59 0.87 0.48 400M 250 DDIM steps, unconditional guidance [32] scale 1.5, 178K train steps, batch size 1200
Table 10. Comparison of a class-conditional ImageNet LDM with recent state-of-the-art methods for class-conditional image generation
on the ImageNet [12] dataset.: Classifier rejection sampling with the given rejection rate as proposed in [67].
D.5. Sample Quality vs. V100 Days (Continued from Sec. 4.1)
Figure 17. For completeness we also report the training progress of class-conditional LDMs on the ImageNet dataset for a fixed number
of 35 V100 days. Results obtained with 100 DDIM steps [84] and = 0. FIDs computed on 5000 samples for efficiency reasons.
For the assessment of sample quality over the training progress in Sec. 4.1, we reported FID and IS scores as a function
of train steps. Another possibility is to report these metrics over the used resources in V100 days. Such an analysis is
additionally provided in Fig. 17, showing qualitatively similar results.
22
Method FID# IS" PSNR" SSIM"
Image Regression [72] 15.2 121.1 27.9 0.801
SR3 [72] 5.2 180.1 26.4 0.762
LDM-4 (ours, 100 steps) 2.8y/4.8z166.3 24.4 3.8 0.690.14
LDM-4 (ours, 50 steps, guiding) 4.4y/6.4z153.7 25.8 3.7 0.740.12
LDM-4 (ours, 100 steps, guiding) 4.4y/6.4z154.1 25.7 3.7 0.730.12
LDM-4 (ours, 100 steps, +15 ep.) 2.6y/4.6z169.76 5.03 24.43.8 0.690.14
Pixel-DM (100 steps, +15 ep.) 5.1y/ 7.1z163.06 4.67 24.13.3 0.590.12
Table 11.4upscaling results on ImageNet-Val. ( 2562);y: FID features computed on validation split,z: FID features computed on train
split. We also include a pixel-space baseline that receives the same amount of compute as LDM-4 . The last two rows received 15 epochs
of additional training compared to the former results.
D.6. Super-Resolution
For better comparability between LDMs and diffusion models in pixel space, we extend our analysis from Tab. 5 by
comparing a diffusion model trained for the same number of steps and with a comparable number1of parameters to our
LDM. The results of this comparison are shown in the last two rows of Tab. 11 and demonstrate that LDM achieves better
performance while allowing for significantly faster sampling. A qualitative comparison is given in Fig. 20 which shows
random samples from both LDM and the diffusion model in pixel space.
D.6.1 LDM-BSR: General Purpose SR Model via Diverse Image Degradation
bicubic LDM-SR LDM-BSR
Figure 18. LDM-BSR generalizes to arbitrary inputs and can be used as a general-purpose upsampler, upscaling samples from a class-
conditional LDM (image cf. Fig. 4) to 10242resolution. In contrast, using a fixed degradation process (see Sec. 4.4) hinders generalization.
To evaluate generalization of our LDM-SR, we apply it both on synthetic LDM samples from a class-conditional ImageNet
model (Sec. 4.1) and images crawled from the internet. Interestingly, we observe that LDM-SR, trained only with a bicubicly
downsampled conditioning as in [72], does not generalize well to images which do not follow this pre-processing. Hence, to
obtain a superresolution model for a wide range of real world images, which can contain complex superpositions of camera
noise, compression artifacts, blurr and interpolations, we replace the bicubic downsampling operation in LDM-SR with the
degration pipeline from [105]. The BSR-degradation process is a degradation pipline which applies JPEG compressions
noise, camera sensor noise, different image interpolations for downsampling, Gaussian blur kernels and Gaussian noise in a
random order to an image. We found that using the bsr-degredation process with the original parameters as in [105] leads to
a very strong degradation process. Since a more moderate degradation process seemed apppropiate for our application, we
adapted the parameters of the bsr-degradation (our adapted degradation process can be found in our code base at https:
//github.com/CompVis/latent-diffusion ). Fig. 18 illustrates the effectiveness of this approach by directly
comparing LDM-SR with LDM-BSR . The latter produces images much sharper than the models confined to a fixed pre-
processing, making it suitable for real-world applications. Further results of LDM-BSR are shown on LSUN-cows in Fig. 19.
1It is not possible to exactly match both architectures since the diffusion model operates in the pixel space
23
E. Implementation Details and Hyperparameters
E.1. Hyperparameters
We provide an overview of the hyperparameters of all trained LDM models in Tab. 12, Tab. 13, Tab. 14 and Tab. 15.
CelebA-HQ 256256 FFHQ 256256 LSUN-Churches 256256 LSUN-Bedrooms 256256
f 4 4 8 4
z-shape 64643 64643 - 64643
jZj 8192 8192 - 8192
Diffusion steps 1000 1000 1000 1000
Noise Schedule linear linear linear linear
Nparams 274M 274M 294M 274M
Channels 224 224 192 224
Depth 2 2 2 2
Channel Multiplier 1,2,3,4 1,2,3,4 1,2,2,4,4 1,2,3,4
Attention resolutions 32, 16, 8 32, 16, 8 32, 16, 8, 4 32, 16, 8
Head Channels 32 32 24 32
Batch Size 48 42 96 48
Iterations410k 635k 500k 1.9M
Learning Rate 9.6e-5 8.4e-5 5.e-5 9.6e-5
Table 12. Hyperparameters for the unconditional LDMs producing the numbers shown in Tab. 1. All models trained on a single NVIDIA
A100.
LDM-1 LDM-2 LDM-4 LDM-8 LDM-16 LDM-32
z-shape 2562563 1281282 64643 32324 16168 88832
jZj - 2048 8192 16384 16384 16384
Diffusion steps 1000 1000 1000 1000 1000 1000
Noise Schedule linear linear linear linear linear linear
Model Size 396M 391M 391M 395M 395M 395M
Channels 192 192 192 256 256 256
Depth 2 2 2 2 2 2
Channel Multiplier 1,1,2,2,4,4 1,2,2,4,4 1,2,3,5 1,2,4 1,2,4 1,2,4
Number of Heads 1 1 1 1 1 1
Batch Size 7 9 40 64 112 112
Iterations 2M 2M 2M 2M 2M 2M
Learning Rate 4.9e-5 6.3e-5 8e-5 6.4e-5 4.5e-5 4.5e-5
Conditioning CA CA CA CA CA CA
CA-resolutions 32, 16, 8 32, 16, 8 32, 16, 8 32, 16, 8 16, 8, 4 8, 4, 2
Embedding Dimension 512 512 512 512 512 512
Transformers Depth 1 1 1 1 1 1
Table 13. Hyperparameters for the conditional LDMs trained on the ImageNet dataset for the analysis in Sec. 4.1. All models trained on a
single NVIDIA A100.
E.2. Implementation Details
E.2.1 Implementations of for conditional LDMs
For the experiments on text-to-image and layout-to-image (Sec. 4.3.1) synthesis, we implement the conditioner as an
unmasked transformer which processes a tokenized version of the input yand produces an output :=(y), where2
RMd. More specifically, the transformer is implemented from Ntransformer blocks consisting of global self-attention
layers, layer-normalization and position-wise MLPs as follows2:
2adapted from https://github.com/lucidrains/x-transformers
24
LDM-1 LDM-2 LDM-4 LDM-8 LDM-16 LDM-32
z-shape 2562563 1281282 64643 32324 16168 88832
jZj - 2048 8192 16384 16384 16384
Diffusion steps 1000 1000 1000 1000 1000 1000
Noise Schedule linear linear linear linear linear linear
Model Size 270M 265M 274M 258M 260M 258M
Channels 192 192 224 256 256 256
Depth 2 2 2 2 2 2
Channel Multiplier 1,1,2,2,4,4 1,2,2,4,4 1,2,3,4 1,2,4 1,2,4 1,2,4
Attention resolutions 32, 16, 8 32, 16, 8 32, 16, 8 32, 16, 8 16, 8, 4 8, 4, 2
Head Channels 32 32 32 32 32 32
Batch Size 9 11 48 96 128 128
Iterations500k 500k 500k 500k 500k 500k
Learning Rate 9e-5 1.1e-4 9.6e-5 9.6e-5 1.3e-4 1.3e-4
Table 14. Hyperparameters for the unconditional LDMs trained on the CelebA dataset for the analysis in Fig. 7. All models trained on a
single NVIDIA A100.: All models are trained for 500k iterations. If converging earlier, we used the best checkpoint for assessing the
provided FID scores.
Task Text-to-Image Layout-to-Image Class-Label-to-Image Super Resolution Inpainting Semantic-Map-to-Image
Dataset LAION OpenImages COCO ImageNet ImageNet Places Landscapes
f 8 4 8 4 4 4 8
z-shape 32324 64643 32324 64643 64643 64643 32324
jZj - 8192 16384 8192 8192 8192 16384
Diffusion steps 1000 1000 1000 1000 1000 1000 1000
Noise Schedule linear linear linear linear linear linear linear
Model Size 1.45B 306M 345M 395M 169M 215M 215M
Channels 320 128 192 192 160 128 128
Depth 2 2 2 2 2 2 2
Channel Multiplier 1,2,4,4 1,2,3,4 1,2,4 1,2,3,5 1,2,2,4 1,4,8 1,4,8
Number of Heads 8 1 1 1 1 1 1
Dropout - - 0.1 - - - -
Batch Size 680 24 48 1200 64 128 48
Iterations 390K 4.4M 170K 178K 860K 360K 360K
Learning Rate 1.0e-4 4.8e-5 4.8e-5 1.0e-4 6.4e-5 1.0e-6 4.8e-5
Conditioning CA CA CA CA concat concat concat
(C)A-resolutions 32, 16, 8 32, 16, 8 32, 16, 8 32, 16, 8 - - -
Embedding Dimension 1280 512 512 512 - - -
Transformer Depth 1 3 2 1 - - -
Table 15. Hyperparameters for the conditional LDMs from Sec. 4. All models trained on a single NVIDIA A100 except for the inpainting
model which was trained on eight V100.
 TokEmb (y) +PosEmb(y) (18)
fori= 1;:::;N :
1 LayerNorm () (19)
2 MultiHeadSelfAttention (1) + (20)
3 LayerNorm (2) (21)
 MLP(3) +2 (22)
 LayerNorm () (23)
(24)
Withavailable, the conditioning is mapped into the UNet via the cross-attention mechanism as depicted in Fig. 3. We
modify the “ablated UNet” [15] architecture and replace the self-attention layer with a shallow (unmasked) transformer
consisting of Tblocks with alternating layers of (i) self-attention, (ii) a position-wise MLP and (iii) a cross-attention layer;
25
see Tab. 16. Note that without (ii) and (iii), this architecture is equivalent to the “ablated UNet”.
While it would be possible to increase the representational power of by additionally conditioning on the time step t, we
do not pursue this choice as it reduces the speed of inference. We leave a more detailed analysis of this modification to future
work.
For the text-to-image model, we rely on a publicly available3tokenizer [99]. The layout-to-image model discretizes the
spatial locations of the bounding boxes and encodes each box as a (l;b;c )-tuple, where ldenotes the (discrete) top-left and b
the bottom-right position. Class information is contained in c.
See Tab. 17 for the hyperparameters of and Tab. 13 for those of the UNet for both of the above tasks.
Note that the class-conditional model as described in Sec. 4.1 is also implemented via cross-attention, where is a single
learnable embedding layer with a dimensionality of 512, mapping classes yto2R1512.
input Rhwc
LayerNorm Rhwc
Conv1x1 Rhwd
nh
Reshape Rh
wd
nh
T8
><
>:SelfAttention
MLP
CrossAttentionRh
wd
nh
Rh
wd
nh
Rh
wd
nh
Reshape Rhwd
nh
Conv1x1 Rhwc
Table 16. Architecture of a transformer block as described in Sec. E.2.1, replacing the self-attention layer of the standard “ablated UNet”
architecture [15]. Here, nhdenotes the number of attention heads and dthe dimensionality per head.
Text-to-Image Layout-to-Image
seq-length 77 92
depthN 32 16
dim 1280 512
Table 17. Hyperparameters for the experiments with transformer encoders in Sec. 4.3.
E.2.2 Inpainting
For our experiments on image-inpainting in Sec. 4.5, we used the code of [88] to generate synthetic masks. We use a fixed
set of 2k validation and 30k testing samples from Places [108]. During training, we use random crops of size 256256
and evaluate on crops of size 512512. This follows the training and testing protocol in [88] and reproduces their reported
metrics (seeyin Tab. 7). We include additional qualitative results of LDM-4, w/ attn in Fig. 21 and of LDM-4, w/o attn, big,
w/ ft in Fig. 22.
E.3. Evaluation Details
This section provides additional details on evaluation for the experiments shown in Sec. 4.
E.3.1 Quantitative Results in Unconditional and Class-Conditional Image Synthesis
We follow common practice and estimate the statistics for calculating the FID-, Precision- and Recall-scores [29,50] shown in
Tab. 1 and 10 based on 50k samples from our models and the entire training set of each of the shown datasets. For calculating
FID scores we use the torch-fidelity package [60]. However, since different data processing pipelines might lead to
different results [64], we also evaluate our models with the script provided by Dhariwal and Nichol [15]. We find that results
3https://huggingface.co/transformers/model_doc/bert.html#berttokenizerfast
26
mainly coincide, except for the ImageNet and LSUN-Bedrooms datasets, where we notice slightly varying scores of 7.76
(torch-fidelity ) vs. 7.77 (Nichol and Dhariwal) and 2.95 vs 3.0. For the future we emphasize the importance of a
unified procedure for sample quality assessment. Precision and Recall are also computed by using the script provided by
Nichol and Dhariwal.
E.3.2 Text-to-Image Synthesis
Following the evaluation protocol of [66] we compute FID and Inception Score for the Text-to-Image models from Tab. 2 by
comparing generated samples with 30000 samples from the validation set of the MS-COCO dataset [51]. FID and Inception
Scores are computed with torch-fidelity .
E.3.3 Layout-to-Image Synthesis
For assessing the sample quality of our Layout-to-Image models from Tab. 9 on the COCO dataset, we follow common
practice [37, 87, 89] and compute FID scores the 2048 unaugmented examples of the COCO Segmentation Challenge split.
To obtain better comparability, we use the exact same samples as in [37]. For the OpenImages dataset we similarly follow
their protocol and use 2048 center-cropped test images from the validation set.
E.3.4 Super Resolution
We evaluate the super-resolution models on ImageNet following the pipeline suggested in [72], i.e. images with a shorter
size less than 256px are removed (both for training and evaluation). On ImageNet, the low-resolution images are produced
using bicubic interpolation with anti-aliasing. FIDs are evaluated using torch-fidelity [60], and we produce samples
on the validation split. For FID scores, we additionally compare to reference features computed on the train split, see Tab. 5
and Tab. 11.
E.3.5 Efficiency Analysis
For efficiency reasons we compute the sample quality metrics plotted in Fig. 6, 17 and 7 based on 5k samples. Therefore,
the results might vary from those shown in Tab. 1 and 10. All models have a comparable number of parameters as provided
in Tab. 13 and 14. We maximize the learning rates of the individual models such that they still train stably. Therefore, the
learning rates slightly vary between different runs cf. Tab. 13 and 14.
E.3.6 User Study
For the results of the user study presented in Tab. 4 we followed the protocoll of [72] and and use the 2-alternative force-choice
paradigm to assess human preference scores for two distinct tasks. In Task-1 subjects were shown a low resolution/masked
image between the corresponding ground truth high resolution/unmasked version and a synthesized image, which was gen-
erated by using the middle image as conditioning. For SuperResolution subjects were asked: ’Which of the two images is a
better high quality version of the low resolution image in the middle?’ . For Inpainting we asked ’Which of the two images
contains more realistic inpainted regions of the image in the middle?’ . In Task-2, humans were similarly shown the low-
res/masked version and asked for preference between two corresponding images generated by the two competing methods.
As in [72] humans viewed the images for 3 seconds before responding.
27
F. Computational Requirements
Method Generator Classifier Overall Inference Nparams FID# IS" Precision"Recall"
Compute Compute Compute Throughput
LSUN Churches 2562
StyleGAN2 [42]y64 - 64 - 59M 3.86 - - -
LDM-8 (ours, 100 steps, 410K) 18 - 18 6.80 256M 4.02 - 0.64 0.52
LSUN Bedrooms 2562
ADM [15]y(1000 steps) 232 - 232 0.03 552M 1.9 - 0.66 0.51
LDM-4 (ours, 200 steps, 1.9M) 60 - 55 1.07 274M 2.95 - 0.66 0.48
CelebA-HQ 2562
LDM-4 (ours, 500 steps, 410K) 14.4 - 14.4 0.43 274M 5.11 - 0.72 0.49
FFHQ 2562
StyleGAN2 [42] 32.13z- 32.13y- 59M 3.8 - - -
LDM-4 (ours, 200 steps, 635K) 26 - 26 1.07 274M 4.98 - 0.73 0.50
ImageNet 2562
VQGAN-f-4 (ours, first stage) 29 - 29 - 55M 0.58yy- - -
VQGAN-f-8 (ours, first stage) 66 - 66 - 68M 1.14yy- - -
BigGAN-deep [3]y128-256 128-256 - 340M 6.95 203.6 2.6 0.87 0.28
ADM [15] (250 steps)y916 - 916 0.12 554M 10.94 100.98 0.69 0.63
ADM-G [15] (25 steps)y916 46 962 0.7 608M 5.58 - 0.81 0.49
ADM-G [15] (250 steps)y916 46 962 0.07 608M 4.59 186.7 0.82 0.52
ADM-G,ADM-U [15] (250 steps)y329 30 349 n/a n/a 3.85 221.72 0.84 0.53
LDM-8-G (ours, 100, 2.9M) 79 12 91 1.93 506M 8.11 190.4 2.6 0.83 0.36
LDM-8 (ours, 200 ddim steps 2.9M, batch size 64) 79 - 79 1.9 395M 17.41 72.92 0.65 0.62
LDM-4 (ours, 250 ddim steps 178K, batch size 1200) 271 - 271 0.7 400M 10.56 103.49 1.24 0.71 0.62
LDM-4-G (ours, 250 ddim steps 178K, batch size 1200, classifier-free guidance [32] scale 1.25) 271 - 271 0.4 400M 3.95 178.22 2.43 0.81 0.55
LDM-4-G (ours, 250 ddim steps 178K, batch size 1200, classifier-free guidance [32] scale 1.5) 271 - 271 0.4 400M 3.60 247.67 5.59 0.87 0.48
Table 18. Comparing compute requirements during training and inference throughput with state-of-the-art generative models. Compute
during training in V100-days, numbers of competing methods taken from [15] unless stated differently;: Throughput measured in sam-
ples/sec on a single NVIDIA A100;y: Numbers taken from [15] ;z: Assumed to be trained on 25M train examples;yy: R-FID vs. ImageNet
validation set
In Tab 18 we provide a more detailed analysis on our used compute ressources and compare our best performing models
on the CelebA-HQ, FFHQ, LSUN and ImageNet datasets with the recent state of the art models by using their provided
numbers, cf. [15]. As they report their used compute in V100 days and we train all our models on a single NVIDIA A100
GPU, we convert the A100 days to V100 days by assuming a 2:2speedup of A100 vs V100 [74]4. To assess sample quality,
we additionally report FID scores on the reported datasets. We closely reach the performance of state of the art methods as
StyleGAN2 [42] and ADM [15] while significantly reducing the required compute resources.
4This factor corresponds to the speedup of the A100 over the V100 for a U-Net, as defined in Fig. 1 in [74]
28
G. Details on Autoencoder Models
We train all our autoencoder models in an adversarial manner following [23], such that a patch-based discriminator D 
is optimized to differentiate original images from reconstructions D(E(x)). To avoid arbitrarily scaled latent spaces, we
regularize the latent zto be zero centered and obtain small variance by introducing an regularizing loss term Lreg.
We investigate two different regularization methods: (i) a low-weighted Kullback-Leibler-term between qE(zjx) =
N(z;E;E2)and a standard normal distribution N(z; 0;1)as in a standard variational autoencoder [46, 69], and, (ii) regu-
larizing the latent space with a vector quantization layer by learning a codebook of jZjdifferent exemplars [96].
To obtain high-fidelity reconstructions we only use a very small regularization for both scenarios, i.e. we either weight the
KLterm by a factor106or choose a high codebook dimensionality jZj.
The full objective to train the autoencoding model (E;D)reads:
LAutoencoder = min
E;Dmax
 
Lrec(x;D(E(x)))Ladv(D(E(x))) + logD (x) +Lreg(x;E;D)
(25)
DM Training in Latent Space Note that for training diffusion models on the learned latent space, we again distinguish two
cases when learning p(z)orp(zjy)(Sec. 4.3): (i) For a KL-regularized latent space, we sample z=E(x)+E(x)
"=:E(x),
where"N(0;1). When rescaling the latent, we estimate the component-wise variance
^2=1
bchwX
b;c;h;w(zb;c;h;w^)2
from the first batch in the data, where ^=1
bchwP
b;c;h;wzb;c;h;w. The output ofEis scaled such that the rescaled latent has
unit standard deviation, i.e.z z
^=E(x)
^. (ii) For a VQ-regularized latent space, we extract zbefore the quantization layer
and absorb the quantization operation into the decoder, i.e. it can be interpreted as the first layer of D.
H. Additional Qualitative Results
Finally, we provide additional qualitative results for our landscapes model (Fig. 12, 23, 24 and 25), our class-conditional
ImageNet model (Fig. 26 - 27) and our unconditional models for the CelebA-HQ, FFHQ and LSUN datasets (Fig. 28 - 31).
Similar as for the inpainting model in Sec. 4.5 we also fine-tuned the semantic landscapes model from Sec. 4.3.2 directly on
5122images and depict qualitative results in Fig. 12 and Fig. 23. For our those models trained on comparably small datasets,
we additionally show nearest neighbors in VGG [79] feature space for samples from our models in Fig. 32 - 34.
29
bicubic LDM-BSR
Figure 19. LDM-BSR generalizes to arbitrary inputs and can be used as a general-purpose upsampler, upscaling samples from the LSUN-
Cows dataset to 10242resolution.
30
input GT Pixel Baseline #1 Pixel Baseline #2 LDM #1 LDM #2
Figure 20. Qualitative superresolution comparison of two random samples between LDM-SR and baseline-diffusionmodel in Pixelspace.
Evaluated on imagenet validation-set after same amount of training steps.
31
input GT LaMa [88] LDM #1 LDM #2 LDM #3
Figure 21. Qualitative results on image inpainting. In contrast to [88], our generative approach enables generation of multiple diverse
samples for a given input.
32
input result input result
Figure 22. More qualitative results on object removal as in Fig. 11.
33
Semantic Synthesis on Flickr-Landscapes [23] ( 5122finetuning)
Figure 23. Convolutional samples from the semantic landscapes model as in Sec. 4.3.2, finetuned on 5122images.
34
Figure 24. A LDM trained on 2562resolution can generalize to larger resolution for spatially conditioned tasks such as semantic synthesis
of landscape images. See Sec. 4.3.2.
35
Semantic Synthesis on Flickr-Landscapes [23]
Figure 25. When provided a semantic map as conditioning, our LDMs generalize to substantially larger resolutions than those seen during
training. Although this model was trained on inputs of size 2562it can be used to create high-resolution samples as the ones shown here,
which are of resolution 1024384. 36
Random class conditional samples on the ImageNet dataset
Figure 26. Random samples from LDM-4 trained on the ImageNet dataset. Sampled with classifier-free guidance [32] scale s= 5:0and
200 DDIM steps with = 1:0.
37
Random class conditional samples on the ImageNet dataset
Figure 27. Random samples from LDM-4 trained on the ImageNet dataset. Sampled with classifier-free guidance [32] scale s= 3:0and
200 DDIM steps with = 1:0.
38
Random samples on the CelebA-HQ dataset
Figure 28. Random samples of our best performing model LDM-4 on the CelebA-HQ dataset. Sampled with 500 DDIM steps and = 0
(FID = 5.15).
39
Random samples on the FFHQ dataset
Figure 29. Random samples of our best performing model LDM-4 on the FFHQ dataset. Sampled with 200 DDIM steps and = 1 (FID
= 4.98).
40
Random samples on the LSUN-Churches dataset
Figure 30. Random samples of our best performing model LDM-8 on the LSUN-Churches dataset. Sampled with 200 DDIM steps and
= 0(FID = 4.48).
41
Random samples on the LSUN-Bedrooms dataset
Figure 31. Random samples of our best performing model LDM-4 on the LSUN-Bedrooms dataset. Sampled with 200 DDIM steps and
= 1(FID = 2.95).
42
Nearest Neighbors on the CelebA-HQ dataset
Figure 32. Nearest neighbors of our best CelebA-HQ model, computed in the feature space of a VGG-16 [79]. The leftmost sample is
from our model. The remaining samples in each row are its 10 nearest neighbors.
43
Nearest Neighbors on the FFHQ dataset
Figure 33. Nearest neighbors of our best FFHQ model, computed in the feature space of a VGG-16 [79]. The leftmost sample is from our
model. The remaining samples in each row are its 10 nearest neighbors.
44
Nearest Neighbors on the LSUN-Churches dataset
Figure 34. Nearest neighbors of our best LSUN-Churches model, computed in the feature space of a VGG-16 [79]. The leftmost sample
is from our model. The remaining samples in each row are its 10 nearest neighbors.
45