<s>
DALL-E	B-General_Concept
(	O
stylized	O
as	O
DALL·E	B-General_Concept
)	O
and	O
DALL-E	B-General_Concept
2	I-General_Concept
are	O
deep	B-Algorithm
learning	I-Algorithm
models	O
developed	O
by	O
OpenAI	O
to	O
generate	O
digital	O
images	O
from	O
natural	O
language	O
descriptions	O
,	O
called	O
"	O
prompts	O
"	O
.	O
</s>
<s>
DALL-E	B-General_Concept
was	O
revealed	O
by	O
OpenAI	O
in	O
a	O
blog	O
post	O
in	O
January	O
2021	O
,	O
and	O
uses	O
a	O
version	O
of	O
GPT-3	B-General_Concept
modified	O
to	O
generate	O
images	O
.	O
</s>
<s>
In	O
April	O
2022	O
,	O
OpenAI	O
announced	O
DALL-E	B-General_Concept
2	I-General_Concept
,	O
a	O
successor	O
designed	O
to	O
generate	O
more	O
realistic	O
images	O
at	O
higher	O
resolutions	O
that	O
"	O
can	O
combine	O
concepts	O
,	O
attributes	O
,	O
and	O
styles	O
"	O
.	O
</s>
<s>
On	O
20	O
July	O
2022	O
,	O
DALL-E	B-General_Concept
2	I-General_Concept
entered	O
into	O
a	O
beta	O
phase	O
with	O
invitations	O
sent	O
to	O
1	O
million	O
waitlisted	O
individuals	O
;	O
users	O
can	O
generate	O
a	O
certain	O
number	O
of	O
images	O
for	O
free	O
every	O
month	O
and	O
may	O
purchase	O
more	O
.	O
</s>
<s>
On	O
28	O
September	O
2022	O
,	O
DALL-E	B-General_Concept
2	I-General_Concept
was	O
opened	O
to	O
anyone	O
and	O
the	O
waitlist	O
requirement	O
was	O
removed	O
.	O
</s>
<s>
In	O
early	O
November	O
2022	O
,	O
OpenAI	O
released	O
DALL-E	B-General_Concept
2	I-General_Concept
as	O
an	O
API	B-General_Concept
,	O
allowing	O
developers	O
to	O
integrate	O
the	O
model	O
into	O
their	O
own	O
applications	O
.	O
</s>
<s>
Microsoft	O
unveiled	O
their	O
implementation	O
of	O
DALL-E	B-General_Concept
2	I-General_Concept
in	O
their	O
Designer	O
app	O
and	O
Image	O
Creator	O
tool	O
included	O
in	O
Bing	B-Application
and	O
Microsoft	B-Application
Edge	I-Application
.	O
</s>
<s>
CALA	O
and	O
Mixtiles	O
are	O
among	O
other	O
early	O
adopters	O
of	O
the	O
DALL-E	B-General_Concept
2	I-General_Concept
API	B-General_Concept
.	O
</s>
<s>
The	O
API	B-General_Concept
operates	O
on	O
a	O
cost	O
per	O
image	O
basis	O
,	O
with	O
prices	O
varying	O
depending	O
on	O
image	O
resolution	O
.	O
</s>
<s>
The	O
Generative	O
Pre-trained	O
Transformer	B-Algorithm
(	O
GPT	O
)	O
model	O
was	O
initially	O
developed	O
by	O
OpenAI	O
in	O
2018	O
,	O
using	O
a	O
Transformer	B-Algorithm
architecture	O
.	O
</s>
<s>
The	O
first	O
iteration	O
,	O
GPT	O
,	O
was	O
scaled	O
up	O
to	O
produce	O
GPT-2	B-General_Concept
in	O
2019	O
;	O
in	O
2020	O
it	O
was	O
scaled	O
up	O
again	O
to	O
produce	O
GPT-3	B-General_Concept
,	O
with	O
175	O
billion	O
parameters	O
.	O
</s>
<s>
DALL-E	B-General_Concept
'	O
s	O
model	O
is	O
a	O
multimodal	O
implementation	O
of	O
GPT-3	B-General_Concept
with	O
12	O
billion	O
parameters	O
which	O
"	O
swaps	O
text	O
for	O
pixels	O
"	O
,	O
trained	O
on	O
text-image	O
pairs	O
from	O
the	O
Internet	O
.	O
</s>
<s>
DALL-E	B-General_Concept
2	I-General_Concept
uses	O
3.5	O
billion	O
parameters	O
,	O
a	O
smaller	O
number	O
than	O
its	O
predecessor	O
.	O
</s>
<s>
DALL-E	B-General_Concept
was	O
developed	O
and	O
announced	O
to	O
the	O
public	O
in	O
conjunction	O
with	O
CLIP	O
(	O
Contrastive	B-General_Concept
Language-Image	I-General_Concept
Pre-training	I-General_Concept
)	O
.	O
</s>
<s>
CLIP	O
is	O
a	O
separate	O
model	O
based	O
on	O
zero-shot	B-Algorithm
learning	I-Algorithm
that	O
was	O
trained	O
on	O
400	O
million	O
pairs	O
of	O
images	O
with	O
text	O
captions	O
scraped	B-Application
from	O
the	O
Internet	O
.	O
</s>
<s>
Its	O
role	O
is	O
to	O
"	O
understand	O
and	O
rank	O
"	O
DALL-E	B-General_Concept
'	O
s	O
output	O
by	O
predicting	O
which	O
caption	O
from	O
a	O
list	O
of	O
32,768	O
captions	O
randomly	O
selected	O
from	O
the	O
dataset	O
(	O
of	O
which	O
one	O
was	O
the	O
correct	O
answer	O
)	O
is	O
most	O
appropriate	O
for	O
an	O
image	O
.	O
</s>
<s>
This	O
model	O
is	O
used	O
to	O
filter	O
a	O
larger	O
initial	O
list	O
of	O
images	O
generated	O
by	O
DALL-E	B-General_Concept
to	O
select	O
the	O
most	O
appropriate	O
outputs	O
.	O
</s>
<s>
DALL-E	B-General_Concept
2	I-General_Concept
uses	O
a	O
diffusion	B-Algorithm
model	I-Algorithm
conditioned	O
on	O
CLIP	O
image	O
embeddings	O
,	O
which	O
,	O
during	O
inference	O
,	O
are	O
generated	O
from	O
CLIP	O
text	O
embeddings	O
by	O
a	O
prior	O
model	O
.	O
</s>
<s>
DALL-E	B-General_Concept
can	O
generate	O
imagery	O
in	O
multiple	O
styles	O
,	O
including	O
photorealistic	O
imagery	O
,	O
paintings	O
,	O
and	O
emoji	B-Application
.	O
</s>
<s>
Thom	O
Dunn	O
writing	O
for	O
BoingBoing	O
remarked	O
that	O
"	O
For	O
example	O
,	O
when	O
asked	O
to	O
draw	O
a	O
daikon	O
radish	O
blowing	O
its	O
nose	O
,	O
sipping	O
a	O
latte	O
,	O
or	O
riding	O
a	O
unicycle	O
,	O
DALL-E	B-General_Concept
often	O
draws	O
the	O
handkerchief	O
,	O
hands	O
,	O
and	O
feet	O
in	O
plausible	O
locations.	O
"	O
</s>
<s>
DALL-E	B-General_Concept
showed	O
the	O
ability	O
to	O
"	O
fill	O
in	O
the	O
blanks	O
"	O
to	O
infer	O
appropriate	O
details	O
without	O
specific	O
prompts	O
,	O
such	O
as	O
adding	O
Christmas	O
imagery	O
to	O
prompts	O
commonly	O
associated	O
with	O
the	O
celebration	O
,	O
and	O
appropriately-placed	O
shadows	O
to	O
images	O
that	O
did	O
not	O
mention	O
them	O
.	O
</s>
<s>
Furthermore	O
,	O
DALL-E	B-General_Concept
exhibits	O
a	O
broad	O
understanding	O
of	O
visual	O
and	O
design	O
trends	O
.	O
</s>
<s>
DALL-E	B-General_Concept
can	O
produce	O
images	O
for	O
a	O
wide	O
variety	O
of	O
arbitrary	O
descriptions	O
from	O
various	O
viewpoints	O
with	O
only	O
rare	O
failures	O
.	O
</s>
<s>
Mark	O
Riedl	O
,	O
an	O
associate	O
professor	O
at	O
the	O
Georgia	O
Tech	O
School	O
of	O
Interactive	O
Computing	O
,	O
found	O
that	O
DALL-E	B-General_Concept
could	O
blend	O
concepts	O
(	O
described	O
as	O
a	O
key	O
element	O
of	O
human	O
creativity	O
)	O
.	O
</s>
<s>
Given	O
an	O
existing	O
image	O
,	O
DALL-E	B-General_Concept
2	I-General_Concept
can	O
produce	O
"	O
variations	O
"	O
of	O
the	O
image	O
as	O
individual	O
outputs	O
based	O
on	O
the	O
original	O
and	O
edit	O
the	O
image	O
to	O
modify	O
or	O
expand	O
upon	O
it	O
.	O
</s>
<s>
DALL-E	B-General_Concept
2	I-General_Concept
's	O
"	O
inpainting	O
"	O
and	O
"	O
outpainting	O
"	O
use	O
context	O
from	O
an	O
image	O
to	O
fill	O
in	O
missing	O
areas	O
using	O
a	O
medium	O
consistent	O
with	O
the	O
original	O
,	O
following	O
a	O
given	O
prompt	O
.	O
</s>
<s>
DALL-E	B-General_Concept
2	I-General_Concept
's	O
reliance	O
on	O
public	O
datasets	O
influences	O
its	O
results	O
and	O
leads	O
to	O
algorithmic	O
bias	O
in	O
some	O
cases	O
,	O
such	O
as	O
generating	O
higher	O
numbers	O
of	O
men	O
than	O
women	O
for	O
requests	O
that	O
do	O
not	O
mention	O
gender	O
.	O
</s>
<s>
DALL-E	B-General_Concept
2	I-General_Concept
's	O
training	O
data	O
was	O
filtered	O
to	O
remove	O
violent	O
and	O
sexual	O
imagery	O
,	O
but	O
this	O
was	O
found	O
to	O
increase	O
bias	O
in	O
some	O
cases	O
such	O
as	O
reducing	O
the	O
frequency	O
of	O
women	O
being	O
generated	O
.	O
</s>
<s>
In	O
September	O
2022	O
,	O
OpenAI	O
confirmed	O
to	O
The	O
Verge	O
that	O
DALL-E	B-General_Concept
invisibly	O
inserts	O
phrases	O
into	O
user	O
prompts	O
to	O
address	O
bias	O
in	O
results	O
;	O
for	O
instance	O
,	O
"	O
black	O
man	O
"	O
and	O
"	O
Asian	O
woman	O
"	O
are	O
inserted	O
into	O
prompts	O
that	O
do	O
not	O
specify	O
gender	O
or	O
race	O
.	O
</s>
<s>
A	O
concern	O
about	O
DALL-E	B-General_Concept
2	I-General_Concept
and	O
similar	O
image	O
generation	O
models	O
is	O
that	O
they	O
could	O
be	O
used	O
to	O
propagate	O
deepfakes	B-Application
and	O
other	O
forms	O
of	O
misinformation	O
.	O
</s>
<s>
Another	O
concern	O
about	O
DALL-E	B-General_Concept
2	I-General_Concept
and	O
similar	O
models	O
is	O
that	O
they	O
could	O
cause	O
technological	O
unemployment	O
for	O
artists	O
,	O
photographers	O
,	O
and	O
graphic	O
designers	O
due	O
to	O
their	O
accuracy	O
and	O
popularity	O
.	O
</s>
<s>
DALL-E	B-General_Concept
2	I-General_Concept
's	O
language	O
understanding	O
has	O
limits	O
.	O
</s>
<s>
Most	O
coverage	O
of	O
DALL-E	B-General_Concept
focuses	O
on	O
a	O
small	O
subset	O
of	O
"	O
surreal	O
"	O
or	O
"	O
quirky	O
"	O
outputs	O
.	O
</s>
<s>
DALL-E	B-General_Concept
'	O
s	O
output	O
for	O
"	O
an	O
illustration	O
of	O
a	O
baby	O
daikon	O
radish	O
in	O
a	O
tutu	O
walking	O
a	O
dog	O
"	O
was	O
mentioned	O
in	O
pieces	O
from	O
Input	O
,	O
NBC	O
,	O
Nature	O
,	O
and	O
other	O
publications	O
.	O
</s>
<s>
ExtremeTech	O
stated	O
"	O
you	O
can	O
ask	O
DALL-E	B-General_Concept
for	O
a	O
picture	O
of	O
a	O
phone	O
or	O
vacuum	O
cleaner	O
from	O
a	O
specified	O
period	O
of	O
time	O
,	O
and	O
it	O
understands	O
how	O
those	O
objects	O
have	O
changed	O
"	O
.	O
</s>
<s>
Engadget	B-Application
also	O
noted	O
its	O
unusual	O
capacity	O
for	O
"	O
understanding	O
how	O
telephones	O
and	O
other	O
objects	O
change	O
over	O
time	O
"	O
.	O
</s>
<s>
According	O
to	O
MIT	O
Technology	O
Review	O
,	O
one	O
of	O
OpenAI	O
's	O
objectives	O
was	O
to	O
"	O
give	O
language	B-Language
models	I-Language
a	O
better	O
grasp	O
of	O
the	O
everyday	O
concepts	O
that	O
humans	O
use	O
to	O
make	O
sense	O
of	O
things	O
"	O
.	O
</s>
<s>
Wall	O
Street	O
investors	O
have	O
had	O
a	O
positive	O
reception	O
of	O
DALL-E	B-General_Concept
2	I-General_Concept
,	O
with	O
some	O
firms	O
thinking	O
it	O
could	O
represent	O
a	O
turning	O
point	O
for	O
a	O
future	O
multi-trillion	O
dollar	O
industry	O
.	O
</s>
<s>
Japan	O
's	O
anime	O
community	O
has	O
had	O
a	O
negative	O
reaction	O
to	O
DALL-E	B-General_Concept
2	I-General_Concept
and	O
similar	O
models	O
.	O
</s>
<s>
AI-driven	O
image	O
generation	O
tools	O
have	O
been	O
heavily	O
criticized	O
by	O
artists	O
because	O
they	O
are	O
trained	O
on	O
human-made	O
art	O
scraped	B-Application
from	O
the	O
web.	O
"	O
</s>
<s>
OpenAI	O
has	O
not	O
released	O
information	O
about	O
what	O
dataset(s )	O
were	O
used	O
to	O
train	O
DALL-E	B-General_Concept
2	I-General_Concept
,	O
inciting	O
concern	O
from	O
some	O
that	O
the	O
work	O
of	O
artists	O
has	O
been	O
used	O
for	O
training	O
without	O
permission	O
.	O
</s>
<s>
There	O
have	O
been	O
several	O
attempts	O
to	O
create	O
open-source	B-License
implementations	O
of	O
DALL-E	B-General_Concept
.	O
</s>
<s>
Released	O
in	O
2022	O
on	O
Hugging	B-Application
Face	I-Application
's	O
Spaces	O
platform	O
,	O
Craiyon	O
(	O
formerly	O
DALL-E	B-General_Concept
Mini	O
until	O
a	O
name	O
change	O
was	O
requested	O
by	O
OpenAI	O
in	O
June	O
2022	O
)	O
is	O
an	O
AI	O
model	O
based	O
on	O
the	O
original	O
DALL-E	B-General_Concept
that	O
was	O
trained	O
on	O
unfiltered	O
data	O
from	O
the	O
Internet	O
.	O
</s>
