<s>
A	O
text-to-image	B-General_Concept
model	I-General_Concept
is	O
a	O
machine	O
learning	O
model	O
which	O
takes	O
as	O
input	O
a	O
natural	O
language	O
description	O
and	O
produces	O
an	O
image	O
matching	O
that	O
description	O
.	O
</s>
<s>
Such	O
models	O
began	O
to	O
be	O
developed	O
in	O
the	O
mid-2010s	O
,	O
as	O
a	O
result	O
of	O
advances	O
in	O
deep	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
.	O
</s>
<s>
In	O
2022	O
,	O
the	O
output	O
of	O
state	O
of	O
the	O
art	O
text-to-image	B-General_Concept
models	O
,	O
such	O
as	O
OpenAI	O
's	O
DALL-E	B-General_Concept
2	I-General_Concept
,	O
Google	O
Brain	O
's	O
Imagen	B-General_Concept
and	O
StabilityAI	O
's	O
Stable	B-General_Concept
Diffusion	I-General_Concept
began	O
to	O
approach	O
the	O
quality	O
of	O
real	O
photographs	O
and	O
human-drawn	O
art	O
.	O
</s>
<s>
Text-to-image	B-General_Concept
models	O
generally	O
combine	O
a	O
language	B-Language
model	I-Language
,	O
which	O
transforms	O
the	O
input	O
text	O
into	O
a	O
latent	O
representation	O
,	O
and	O
a	O
generative	O
image	O
model	O
,	O
which	O
produces	O
an	O
image	O
conditioned	O
on	O
that	O
representation	O
.	O
</s>
<s>
Before	O
the	O
rise	O
of	O
deep	B-Algorithm
learning	I-Algorithm
,	O
attempts	O
to	O
build	O
text-to-image	B-General_Concept
models	O
were	O
limited	O
to	O
collages	O
by	O
arranging	O
existing	O
component	O
images	O
,	O
such	O
as	O
from	O
a	O
database	O
of	O
clip	O
art	O
.	O
</s>
<s>
The	O
inverse	O
task	O
,	O
image	O
captioning	O
,	O
was	O
more	O
tractable	O
and	O
a	O
number	O
of	O
image	O
captioning	O
deep	B-Algorithm
learning	I-Algorithm
models	O
came	O
prior	O
to	O
the	O
first	O
text-to-image	B-General_Concept
models	O
.	O
</s>
<s>
The	O
first	O
modern	O
text-to-image	B-General_Concept
model	I-General_Concept
,	O
alignDRAW	O
,	O
was	O
introduced	O
in	O
2015	O
by	O
researchers	O
from	O
the	O
University	O
of	O
Toronto	O
.	O
</s>
<s>
alignDRAW	O
extended	O
the	O
previously-introduced	O
DRAW	O
architecture	O
(	O
which	O
used	O
a	O
recurrent	B-Algorithm
variational	B-Algorithm
autoencoder	I-Algorithm
with	O
an	O
attention	B-General_Concept
mechanism	I-General_Concept
)	O
to	O
be	O
conditioned	O
on	O
text	O
sequences	O
.	O
</s>
<s>
became	O
the	O
first	O
to	O
use	O
generative	B-Algorithm
adversarial	I-Algorithm
networks	I-Algorithm
for	O
the	O
text-to-image	B-General_Concept
task	O
.	O
</s>
<s>
One	O
of	O
the	O
first	O
text-to-image	B-General_Concept
models	O
to	O
capture	O
widespread	O
public	O
attention	O
was	O
OpenAI	O
's	O
DALL-E	B-General_Concept
,	O
a	O
transformer	B-Algorithm
system	O
announced	O
in	O
January	O
2021	O
.	O
</s>
<s>
A	O
successor	O
capable	O
of	O
generating	O
more	O
complex	O
and	O
realistic	O
images	O
,	O
DALL-E	B-General_Concept
2	I-General_Concept
,	O
was	O
unveiled	O
in	O
April	O
2022	O
,	O
followed	O
by	O
Stable	B-General_Concept
Diffusion	I-General_Concept
publicly	O
released	O
in	O
August	O
2022	O
.	O
</s>
<s>
Following	O
other	O
text-to-image	B-General_Concept
models	O
,	O
language	O
model-powered	O
text-to-video	B-General_Concept
platforms	O
such	O
as	O
Runway	O
,	O
Make-A-Video	O
,	O
Imagen	B-General_Concept
Video	O
,	O
Midjourney	O
,	O
and	O
Phenaki	O
can	O
generate	O
video	O
from	O
text	O
and/or	O
text/image	O
prompts	O
.	O
</s>
<s>
Text-to-image	B-General_Concept
models	O
have	O
been	O
built	O
using	O
a	O
variety	O
of	O
architectures	O
.	O
</s>
<s>
The	O
text	O
encoding	O
step	O
may	O
be	O
performed	O
with	O
a	O
recurrent	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
such	O
as	O
a	O
long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
(	O
LSTM	B-Algorithm
)	O
network	O
,	O
though	O
transformer	B-Algorithm
models	I-Algorithm
have	O
since	O
become	O
a	O
more	O
popular	O
option	O
.	O
</s>
<s>
For	O
the	O
image	O
generation	O
step	O
,	O
conditional	O
generative	B-Algorithm
adversarial	I-Algorithm
networks	I-Algorithm
have	O
been	O
commonly	O
used	O
,	O
with	O
diffusion	B-Algorithm
models	I-Algorithm
also	O
becoming	O
a	O
popular	O
option	O
in	O
recent	O
years	O
.	O
</s>
<s>
Rather	O
than	O
directly	O
training	O
a	O
model	O
to	O
output	O
a	O
high-resolution	O
image	O
conditioned	O
on	O
a	O
text	O
embedding	O
,	O
a	O
popular	O
technique	O
is	O
to	O
train	O
a	O
model	O
to	O
generate	O
low-resolution	O
images	O
,	O
and	O
use	O
one	O
or	O
more	O
auxiliary	O
deep	B-Algorithm
learning	I-Algorithm
models	O
to	O
upscale	O
it	O
,	O
filling	O
in	O
finer	O
details	O
.	O
</s>
<s>
Text-to-image	B-General_Concept
models	O
are	O
trained	O
on	O
large	O
datasets	O
of	O
(	O
text	O
,	O
image	O
)	O
pairs	O
,	O
often	O
scraped	O
from	O
the	O
web	O
.	O
</s>
<s>
With	O
their	O
2022	O
Imagen	B-General_Concept
model	O
,	O
Google	O
Brain	O
reported	O
positive	O
results	O
from	O
using	O
a	O
large	O
language	B-Language
model	I-Language
trained	O
separately	O
on	O
a	O
text-only	O
corpus	O
(	O
with	O
its	O
weights	O
subsequently	O
frozen	O
)	O
,	O
a	O
departure	O
from	O
the	O
theretofore	O
standard	O
approach	O
.	O
</s>
<s>
Training	O
a	O
text-to-image	B-General_Concept
model	I-General_Concept
requires	O
a	O
dataset	O
of	O
images	O
paired	O
with	O
text	O
captions	O
.	O
</s>
<s>
It	O
is	O
considered	O
less	O
difficult	O
to	O
train	O
a	O
high-quality	O
text-to-image	B-General_Concept
model	I-General_Concept
with	O
these	O
datasets	O
,	O
because	O
of	O
their	O
narrow	O
range	O
of	O
subject	O
matter	O
.	O
</s>
<s>
Evaluating	O
and	O
comparing	O
the	O
quality	O
of	O
text-to-image	B-General_Concept
models	O
is	O
a	O
challenging	O
problem	O
,	O
and	O
involves	O
assessing	O
multiple	O
desirable	O
properties	O
.	O
</s>
<s>
A	O
desideratum	O
specific	O
to	O
text-to-image	B-General_Concept
models	O
is	O
that	O
generated	O
images	O
semantically	O
align	O
with	O
the	O
text	O
captions	O
used	O
to	O
generate	O
them	O
.	O
</s>
<s>
A	O
common	O
algorithmic	O
metric	O
for	O
assessing	O
image	O
quality	O
and	O
diversity	O
is	O
Inception	B-General_Concept
score	I-General_Concept
(	O
IS	O
)	O
,	O
which	O
is	O
based	O
on	O
the	O
distribution	O
of	O
labels	O
predicted	O
by	O
a	O
pretrained	O
Inceptionv3	B-General_Concept
image	O
classification	O
model	O
when	O
applied	O
to	O
a	O
sample	O
of	O
images	O
generated	O
by	O
the	O
text-to-image	B-General_Concept
model	I-General_Concept
.	O
</s>
<s>
Another	O
popular	O
metric	O
is	O
the	O
related	O
Fréchet	B-Algorithm
inception	I-Algorithm
distance	I-Algorithm
,	O
which	O
compares	O
the	O
distribution	O
of	O
generated	O
images	O
and	O
real	O
training	O
images	O
,	O
according	O
to	O
features	O
extracted	O
by	O
one	O
of	O
the	O
final	O
layers	O
of	O
a	O
pretrained	O
image	O
classification	O
model	O
.	O
</s>
