<s>
Stable	B-General_Concept
Diffusion	I-General_Concept
is	O
a	O
deep	B-Algorithm
learning	I-Algorithm
,	O
text-to-image	B-General_Concept
model	I-General_Concept
released	O
in	O
2022	O
.	O
</s>
<s>
It	O
is	O
primarily	O
used	O
to	O
generate	O
detailed	O
images	O
conditioned	O
on	O
text	O
descriptions	O
,	O
though	O
it	O
can	O
also	O
be	O
applied	O
to	O
other	O
tasks	O
such	O
as	O
inpainting	B-Algorithm
,	O
outpainting	O
,	O
and	O
generating	O
image-to-image	O
translations	O
guided	O
by	O
a	O
text	B-General_Concept
prompt	I-General_Concept
.	O
</s>
<s>
It	O
was	O
developed	O
by	O
the	O
start-up	O
Stability	B-General_Concept
AI	I-General_Concept
in	O
collaboration	O
with	O
a	O
number	O
of	O
academic	O
researchers	O
and	O
non-profit	O
organizations	O
.	O
</s>
<s>
Stable	B-General_Concept
Diffusion	I-General_Concept
is	O
a	O
latent	O
diffusion	B-Algorithm
model	I-Algorithm
,	O
a	O
kind	O
of	O
deep	O
generative	O
neural	B-Architecture
network	I-Architecture
.	O
</s>
<s>
Its	O
code	O
and	O
model	O
weights	O
have	O
been	O
released	O
publicly	B-License
,	O
and	O
it	O
can	O
run	O
on	O
most	O
consumer	O
hardware	O
equipped	O
with	O
a	O
modest	O
GPU	B-Architecture
with	O
at	O
least	O
8GB	O
VRAM	O
.	O
</s>
<s>
This	O
marked	O
a	O
departure	O
from	O
previous	O
proprietary	O
text-to-image	B-General_Concept
models	O
such	O
as	O
DALL-E	B-General_Concept
and	O
Midjourney	B-Application
which	O
were	O
accessible	O
only	O
via	O
cloud	B-Architecture
services	I-Architecture
.	O
</s>
<s>
The	O
development	O
of	O
Stable	B-General_Concept
Diffusion	I-General_Concept
was	O
funded	O
and	O
shaped	O
by	O
the	O
start-up	O
company	O
Stability	B-General_Concept
AI	I-General_Concept
.	O
</s>
<s>
Development	O
was	O
led	O
by	O
Patrick	O
Esser	O
of	O
Runway	O
and	O
Robin	O
Rombach	O
of	O
CompVis	O
,	O
who	O
were	O
among	O
the	O
researchers	O
who	O
had	O
earlier	O
invented	O
the	O
latent	O
diffusion	B-Algorithm
model	I-Algorithm
architecture	O
used	O
by	O
Stable	B-General_Concept
Diffusion	I-General_Concept
.	O
</s>
<s>
Stability	B-General_Concept
AI	I-General_Concept
also	O
credited	O
Eleuther	O
AI	O
and	O
LAION	O
(	O
a	O
German	O
nonprofit	O
which	O
assembled	O
the	O
dataset	O
on	O
which	O
Stable	B-General_Concept
Diffusion	I-General_Concept
was	O
trained	O
)	O
as	O
supporters	O
of	O
the	O
project	O
.	O
</s>
<s>
In	O
October	O
2022	O
,	O
Stability	B-General_Concept
AI	I-General_Concept
raised	O
US$101	O
million	O
in	O
a	O
round	O
led	O
by	O
Lightspeed	O
Venture	O
Partners	O
and	O
Coatue	O
Management	O
.	O
</s>
<s>
Stable	B-General_Concept
Diffusion	I-General_Concept
uses	O
a	O
kind	O
of	O
diffusion	B-Algorithm
model	I-Algorithm
(	O
DM	O
)	O
,	O
called	O
a	O
latent	O
diffusion	B-Algorithm
model	I-Algorithm
(	O
LDM	O
)	O
developed	O
by	O
the	O
CompVis	O
group	O
at	O
LMU	O
Munich	O
.	O
</s>
<s>
Introduced	O
in	O
2015	O
,	O
diffusion	B-Algorithm
models	I-Algorithm
are	O
trained	O
with	O
the	O
objective	O
of	O
removing	O
successive	O
applications	O
of	O
Gaussian	O
noise	O
on	O
training	O
images	O
which	O
can	O
be	O
thought	O
of	O
as	O
a	O
sequence	O
of	O
denoising	O
autoencoders	O
.	O
</s>
<s>
Stable	B-General_Concept
Diffusion	I-General_Concept
consists	O
of	O
3	O
parts	O
:	O
the	O
variational	B-Algorithm
autoencoder	I-Algorithm
(	O
VAE	O
)	O
,	O
U-Net	B-Application
,	O
and	O
an	O
optional	O
text	O
encoder	O
.	O
</s>
<s>
The	O
VAE	O
encoder	O
compresses	O
the	O
image	O
from	O
pixel	O
space	O
to	O
a	O
smaller	O
dimensional	O
latent	B-Algorithm
space	I-Algorithm
,	O
capturing	O
a	O
more	O
fundamental	O
semantic	O
meaning	O
of	O
the	O
image	O
.	O
</s>
<s>
The	O
U-Net	B-Application
block	O
,	O
composed	O
of	O
a	O
ResNet	B-Algorithm
backbone	O
,	O
denoises	O
the	O
output	O
from	O
forward	O
diffusion	O
backwards	O
to	O
obtain	O
a	O
latent	O
representation	O
.	O
</s>
<s>
The	O
encoded	O
conditioning	O
data	O
is	O
exposed	O
to	O
denoising	O
U-Nets	B-Application
via	O
a	O
cross-attention	B-General_Concept
mechanism	I-General_Concept
.	O
</s>
<s>
For	O
conditioning	O
on	O
text	O
,	O
the	O
fixed	O
,	O
pretrained	O
CLIP	O
ViT-L/14	O
text	O
encoder	O
is	O
used	O
to	O
transform	O
text	O
prompts	O
to	O
an	O
embedding	B-Algorithm
space	I-Algorithm
.	O
</s>
<s>
Stable	B-General_Concept
Diffusion	I-General_Concept
was	O
trained	O
on	O
pairs	O
of	O
images	O
and	O
captions	O
taken	O
from	O
LAION-5B	O
,	O
a	O
publicly	B-License
available	O
dataset	O
derived	O
from	O
Common	O
Crawl	O
data	O
scraped	O
from	O
the	O
web	O
,	O
where	O
5	O
billion	O
image-text	O
pairs	O
were	O
classified	O
based	O
on	O
language	O
and	O
filtered	O
into	O
separate	O
datasets	O
by	O
resolution	O
,	O
a	O
predicted	O
likelihood	O
of	O
containing	O
a	O
watermark	B-General_Concept
,	O
and	O
predicted	O
"	O
aesthetic	O
"	O
score	O
(	O
e.g.	O
</s>
<s>
The	O
dataset	O
was	O
created	O
by	O
LAION	O
,	O
a	O
German	O
non-profit	O
which	O
receives	O
funding	O
from	O
Stability	B-General_Concept
AI	I-General_Concept
.	O
</s>
<s>
The	O
Stable	B-General_Concept
Diffusion	I-General_Concept
model	O
was	O
trained	O
on	O
three	O
subsets	O
of	O
LAION-5B	O
:	O
laion2B-en	O
,	O
laion-high-resolution	O
,	O
and	O
laion-aesthetics	O
v2	O
5+	O
.	O
</s>
<s>
A	O
third-party	O
analysis	O
of	O
the	O
model	O
's	O
training	O
data	O
identified	O
that	O
out	O
of	O
a	O
smaller	O
subset	O
of	O
12	O
million	O
images	O
taken	O
from	O
the	O
original	O
wider	O
dataset	O
used	O
,	O
approximately	O
47%	O
of	O
the	O
sample	O
size	O
of	O
images	O
came	O
from	O
100	O
different	O
domains	O
,	O
with	O
Pinterest	B-Application
taking	O
up	O
8.5	O
%	O
of	O
the	O
subset	O
,	O
followed	O
by	O
websites	O
such	O
as	O
WordPress	B-Application
,	O
Blogspot	B-Application
,	O
Flickr	B-Algorithm
,	O
DeviantArt	O
and	O
Wikimedia	O
Commons	O
.	O
</s>
<s>
The	O
LAION-Aesthetics	O
v2	O
5+	O
subset	O
also	O
excluded	O
low-resolution	O
images	O
and	O
images	O
which	O
LAION-5B-WatermarkDetection	O
identified	O
as	O
carrying	O
a	O
watermark	B-General_Concept
with	O
greater	O
than	O
80%	O
probability	O
.	O
</s>
<s>
The	O
model	O
was	O
trained	O
using	O
256	O
Nvidia	B-General_Concept
A100	I-General_Concept
GPUs	B-Architecture
on	O
Amazon	B-Application
Web	I-Application
Services	I-Application
for	O
a	O
total	O
of	O
150,000	O
GPU-hours	O
,	O
at	O
a	O
cost	O
of	O
$	O
600,000	O
.	O
</s>
<s>
Stable	B-General_Concept
Diffusion	I-General_Concept
has	O
issues	O
with	O
degradation	O
and	O
inaccuracies	O
in	O
certain	O
scenarios	O
.	O
</s>
<s>
Initial	O
releases	O
of	O
the	O
model	O
were	O
trained	O
on	O
a	O
dataset	O
that	O
consists	O
of	O
512×512	O
resolution	O
images	O
,	O
meaning	O
that	O
the	O
quality	O
of	O
generated	O
images	O
noticeably	O
degrades	O
when	O
user	O
specifications	O
deviate	O
from	O
its	O
"	O
expected	O
"	O
512×512	O
resolution	O
;	O
the	O
version	O
2.0	O
update	O
of	O
the	O
Stable	B-General_Concept
Diffusion	I-General_Concept
model	O
later	O
introduced	O
the	O
ability	O
to	O
natively	O
generate	O
images	O
at	O
768×768	O
resolution	O
.	O
</s>
<s>
Fine-tuned	O
adaptations	O
of	O
Stable	B-General_Concept
Diffusion	I-General_Concept
created	O
through	O
additional	O
retraining	O
have	O
been	O
used	O
for	O
a	O
variety	O
of	O
different	O
use-cases	O
,	O
from	O
medical	O
imaging	O
to	O
algorithmically-generated	B-Application
music	I-Application
.	O
</s>
<s>
For	O
example	O
,	O
the	O
training	O
process	O
for	O
waifu-diffusion	O
requires	O
a	O
minimum	O
30GB	O
of	O
VRAM	O
,	O
which	O
exceeds	O
the	O
usual	O
resource	O
provided	O
in	O
consumer	O
GPUs	B-Architecture
,	O
such	O
as	O
Nvidia	O
’s	O
GeForce	B-Device
30	I-Device
series	I-Device
having	O
around	O
12	O
GB	O
.	O
</s>
<s>
The	O
creators	O
of	O
Stable	B-General_Concept
Diffusion	I-General_Concept
acknowledge	O
the	O
potential	O
for	O
algorithmic	O
bias	O
,	O
as	O
the	O
model	O
was	O
primarily	O
trained	O
on	O
images	O
with	O
English	O
descriptions	O
.	O
</s>
<s>
There	O
are	O
three	O
methods	O
in	O
which	O
user-accessible	O
fine-tuning	O
can	O
be	O
applied	O
to	O
a	O
Stable	B-General_Concept
Diffusion	I-General_Concept
model	O
checkpoint	O
:	O
</s>
<s>
A	O
"	O
hypernetwork	O
"	O
is	O
a	O
small	O
pre-trained	O
neural	B-Architecture
network	I-Architecture
that	O
is	O
applied	O
to	O
various	O
points	O
within	O
a	O
larger	O
neural	B-Architecture
network	I-Architecture
,	O
and	O
refers	O
to	O
the	O
technique	O
created	O
by	O
NovelAI	B-Application
developer	O
Kurumuz	O
in	O
2021	O
,	O
originally	O
intended	O
for	O
text-generation	O
transformer	B-Algorithm
models	I-Algorithm
.	O
</s>
<s>
Hypernetworks	O
steer	O
results	O
towards	O
a	O
particular	O
direction	O
,	O
allowing	O
Stable	O
Diffusion-based	O
models	O
to	O
imitate	O
the	O
art	O
style	O
of	O
specific	O
artists	O
,	O
even	O
if	O
the	O
artist	O
is	O
not	O
recognised	O
by	O
the	O
original	O
model	O
;	O
they	O
process	O
the	O
image	O
by	O
finding	O
key	O
areas	O
of	O
importance	O
such	O
as	O
hair	O
and	O
eyes	O
,	O
and	O
then	O
patch	O
these	O
areas	O
in	O
secondary	O
latent	B-Algorithm
space	I-Algorithm
.	O
</s>
<s>
DreamBooth	B-General_Concept
is	O
a	O
deep	B-Algorithm
learning	I-Algorithm
generation	O
model	O
developed	O
by	O
researchers	O
from	O
Google	B-Application
Research	I-Application
and	O
Boston	O
University	O
in	O
2022	O
which	O
can	O
fine-tune	O
the	O
model	O
to	O
generate	O
precise	O
,	O
personalised	O
outputs	O
that	O
depict	O
a	O
specific	O
subject	O
,	O
following	O
training	O
via	O
a	O
set	O
of	O
images	O
which	O
depict	O
the	O
subject	O
.	O
</s>
<s>
The	O
Stable	B-General_Concept
Diffusion	I-General_Concept
model	O
supports	O
the	O
ability	O
to	O
generate	O
new	O
images	O
from	O
scratch	O
through	O
the	O
use	O
of	O
a	O
text	B-General_Concept
prompt	I-General_Concept
describing	O
elements	O
to	O
be	O
included	O
or	O
omitted	O
from	O
the	O
output	O
.	O
</s>
<s>
Existing	O
images	O
can	O
be	O
re-drawn	O
by	O
the	O
model	O
to	O
incorporate	O
new	O
elements	O
described	O
by	O
a	O
text	B-General_Concept
prompt	I-General_Concept
(	O
a	O
process	O
known	O
as	O
"	O
guided	O
image	O
synthesis	O
"	O
)	O
through	O
its	O
diffusion-denoising	O
mechanism	O
.	O
</s>
<s>
In	O
addition	O
,	O
the	O
model	O
also	O
allows	O
the	O
use	O
of	O
prompts	O
to	O
partially	O
alter	O
existing	O
images	O
via	O
inpainting	B-Algorithm
and	O
outpainting	O
,	O
when	O
used	O
with	O
an	O
appropriate	O
user	O
interface	O
that	O
supports	O
such	O
features	O
,	O
of	O
which	O
numerous	O
different	O
open	O
source	O
implementations	O
exist	O
.	O
</s>
<s>
Stable	B-General_Concept
Diffusion	I-General_Concept
is	O
recommended	O
to	O
be	O
run	O
with	O
10GB	O
or	O
more	O
VRAM	O
,	O
however	O
users	O
with	O
less	O
VRAM	O
may	O
opt	O
to	O
load	O
the	O
weights	O
in	O
float16	O
precision	O
instead	O
of	O
the	O
default	O
float32	O
to	O
tradeoff	O
model	O
performance	O
with	O
lower	O
VRAM	O
usage	O
.	O
</s>
<s>
The	O
text	O
to	O
image	O
sampling	O
script	O
within	O
Stable	B-General_Concept
Diffusion	I-General_Concept
,	O
known	O
as	O
"	O
txt2img	O
"	O
,	O
consumes	O
a	O
text	B-General_Concept
prompt	I-General_Concept
in	O
addition	O
to	O
assorted	O
option	O
parameters	O
covering	O
sampling	O
types	O
,	O
output	O
image	O
dimensions	O
,	O
and	O
seed	O
values	O
.	O
</s>
<s>
Generated	O
images	O
are	O
tagged	O
with	O
an	O
invisible	O
digital	O
watermark	B-General_Concept
to	O
allow	O
users	O
to	O
identify	O
an	O
image	O
as	O
generated	O
by	O
Stable	B-General_Concept
Diffusion	I-General_Concept
,	O
although	O
this	O
watermark	B-General_Concept
loses	O
its	O
efficacy	O
if	O
the	O
image	O
is	O
resized	O
or	O
rotated	O
.	O
</s>
<s>
Each	O
txt2img	O
generation	O
will	O
involve	O
a	O
specific	O
seed	B-Algorithm
value	I-Algorithm
which	O
affects	O
the	O
output	O
image	O
.	O
</s>
<s>
Additional	O
text2img	O
features	O
are	O
provided	O
by	O
front-end	B-Architecture
implementations	O
of	O
Stable	B-General_Concept
Diffusion	I-General_Concept
,	O
which	O
allow	O
users	O
to	O
modify	O
the	O
weight	O
given	O
to	O
specific	O
parts	O
of	O
the	O
text	B-General_Concept
prompt	I-General_Concept
.	O
</s>
<s>
Negative	O
prompts	O
are	O
a	O
feature	O
included	O
in	O
some	O
front-end	B-Architecture
implementations	O
,	O
including	O
Stability	B-General_Concept
AI	I-General_Concept
's	O
own	O
DreamStudio	O
cloud	B-Architecture
service	I-Architecture
,	O
and	O
allow	O
the	O
user	O
to	O
specify	O
prompts	O
which	O
the	O
model	O
should	O
avoid	O
during	O
image	O
generation	O
.	O
</s>
<s>
Stable	B-General_Concept
Diffusion	I-General_Concept
also	O
includes	O
another	O
sampling	O
script	O
,	O
"	O
img2img	O
"	O
,	O
which	O
consumes	O
a	O
text	B-General_Concept
prompt	I-General_Concept
,	O
path	O
to	O
an	O
existing	O
image	O
,	O
and	O
strength	O
value	O
between	O
0.0	O
and	O
1.0	O
.	O
</s>
<s>
The	O
script	O
outputs	O
a	O
new	O
image	O
based	O
on	O
the	O
original	O
image	O
that	O
also	O
features	O
elements	O
provided	O
within	O
the	O
text	B-General_Concept
prompt	I-General_Concept
.	O
</s>
<s>
The	O
ability	O
of	O
img2img	O
to	O
add	O
noise	O
to	O
the	O
original	O
image	O
makes	O
it	O
potentially	O
useful	O
for	O
data	O
anonymization	O
and	O
data	B-General_Concept
augmentation	I-General_Concept
,	O
in	O
which	O
the	O
visual	O
features	O
of	O
image	O
data	O
are	O
changed	O
and	O
anonymized	O
.	O
</s>
<s>
Additionally	O
,	O
Stable	B-General_Concept
Diffusion	I-General_Concept
has	O
been	O
experimented	O
with	O
as	O
a	O
tool	O
for	O
image	O
compression	O
.	O
</s>
<s>
Compared	O
to	O
JPEG	O
and	O
WebP	O
,	O
the	O
recent	O
methods	O
used	O
for	O
image	O
compression	O
in	O
Stable	B-General_Concept
Diffusion	I-General_Concept
face	O
limitations	O
in	O
preserving	O
small	O
text	O
and	O
faces	O
.	O
</s>
<s>
Additional	O
use-cases	O
for	O
image	O
modification	O
via	O
img2img	O
are	O
offered	O
by	O
numerous	O
front-end	B-Architecture
implementations	O
of	O
the	O
Stable	B-General_Concept
Diffusion	I-General_Concept
model	O
.	O
</s>
<s>
Inpainting	B-Algorithm
involves	O
selectively	O
modifying	O
a	O
portion	O
of	O
an	O
existing	O
image	O
delineated	O
by	O
a	O
user-provided	O
layer	O
mask	O
,	O
which	O
fills	O
the	O
masked	O
space	O
with	O
newly	O
generated	O
content	O
based	O
on	O
the	O
provided	O
prompt	O
.	O
</s>
<s>
A	O
dedicated	O
model	O
specifically	O
fine-tuned	O
for	O
inpainting	B-Algorithm
use-cases	O
was	O
created	O
by	O
Stability	B-General_Concept
AI	I-General_Concept
alongside	O
the	O
release	O
of	O
Stable	B-General_Concept
Diffusion	I-General_Concept
2.0	O
.	O
</s>
<s>
A	O
depth-guided	O
model	O
,	O
named	O
"	O
depth2img	O
"	O
,	O
was	O
introduced	O
with	O
the	O
release	O
of	O
Stable	B-General_Concept
Diffusion	I-General_Concept
2.0	O
on	O
November	O
24	O
,	O
2022	O
;	O
this	O
model	O
infers	O
the	O
depth	O
of	O
the	O
provided	O
input	O
image	O
,	O
and	O
generates	O
a	O
new	O
output	O
image	O
based	O
on	O
both	O
the	O
text	B-General_Concept
prompt	I-General_Concept
and	O
the	O
depth	O
information	O
,	O
which	O
allows	O
the	O
coherence	O
and	O
depth	O
of	O
the	O
original	O
input	O
image	O
to	O
be	O
maintained	O
in	O
the	O
generated	O
output	O
.	O
</s>
<s>
ControlNet	O
is	O
a	O
neural	B-Architecture
network	I-Architecture
architecture	O
designed	O
to	O
manage	O
diffusion	B-Algorithm
models	I-Algorithm
by	O
incorporating	O
additional	O
conditions	O
.	O
</s>
<s>
It	O
duplicates	O
the	O
weights	O
of	O
neural	B-Architecture
network	I-Architecture
blocks	O
into	O
a	O
"	O
locked	O
"	O
copy	O
and	O
a	O
"	O
trainable	O
"	O
copy	O
.	O
</s>
<s>
This	O
approach	O
ensures	O
that	O
training	O
with	O
small	O
datasets	O
of	O
image	O
pairs	O
does	O
not	O
compromise	O
the	O
integrity	O
of	O
production-ready	O
diffusion	B-Algorithm
models	I-Algorithm
.	O
</s>
<s>
Stable	B-General_Concept
Diffusion	I-General_Concept
claims	O
no	O
rights	O
on	O
generated	O
images	O
and	O
freely	O
gives	O
users	O
the	O
rights	O
of	O
usage	O
to	O
any	O
generated	O
images	O
from	O
the	O
model	O
provided	O
that	O
the	O
image	O
content	O
is	O
not	O
illegal	O
or	O
harmful	O
to	O
individuals	O
.	O
</s>
<s>
The	O
freedom	O
provided	O
to	O
users	O
over	O
image	O
usage	O
has	O
caused	O
controversy	O
over	O
the	O
ethics	O
of	O
ownership	O
,	O
as	O
Stable	B-General_Concept
Diffusion	I-General_Concept
and	O
other	O
generative	O
models	O
are	O
trained	O
from	O
copyrighted	O
images	O
without	O
the	O
owner	O
’s	O
consent	O
.	O
</s>
<s>
As	O
visual	O
styles	O
and	O
compositions	O
are	O
not	O
subject	O
to	O
copyright	O
,	O
it	O
is	O
often	O
interpreted	O
that	O
users	O
of	O
Stable	B-General_Concept
Diffusion	I-General_Concept
who	O
generate	O
images	O
of	O
artworks	O
should	O
not	O
be	O
considered	O
to	O
be	O
infringing	O
upon	O
the	O
copyright	O
of	O
visually	O
similar	O
works	O
.	O
</s>
<s>
Nonetheless	O
,	O
visual	O
artists	O
have	O
expressed	O
concern	O
that	O
widespread	O
usage	O
of	O
image	O
synthesis	O
software	O
such	O
as	O
Stable	B-General_Concept
Diffusion	I-General_Concept
may	O
eventually	O
lead	O
to	O
human	O
artists	O
,	O
along	O
with	O
photographers	O
,	O
models	O
,	O
cinematographers	O
,	O
and	O
actors	O
,	O
gradually	O
losing	O
commercial	O
viability	O
against	O
AI-based	O
competitors	O
.	O
</s>
<s>
Stable	B-General_Concept
Diffusion	I-General_Concept
is	O
notably	O
more	O
permissive	O
in	O
the	O
types	O
of	O
content	O
users	O
may	O
generate	O
,	O
such	O
as	O
violent	O
or	O
sexually	O
explicit	O
imagery	O
,	O
in	O
comparison	O
to	O
other	O
commercial	O
products	O
based	O
on	O
generative	O
AI	O
.	O
</s>
<s>
Addressing	O
the	O
concerns	O
that	O
the	O
model	O
may	O
be	O
used	O
for	O
abusive	O
purposes	O
,	O
CEO	O
of	O
Stability	B-General_Concept
AI	I-General_Concept
,	O
Emad	O
Mostaque	O
,	O
explains	O
that	O
"	O
[	O
it	O
is ]	O
peoples	O
 '	O
responsibility	O
as	O
to	O
whether	O
they	O
are	O
ethical	O
,	O
moral	O
,	O
and	O
legal	O
in	O
how	O
they	O
operate	O
this	O
technology	O
"	O
,	O
and	O
that	O
putting	O
the	O
capabilities	O
of	O
Stable	B-General_Concept
Diffusion	I-General_Concept
into	O
the	O
hands	O
of	O
the	O
public	O
would	O
result	O
in	O
the	O
technology	O
providing	O
a	O
net	O
benefit	O
,	O
in	O
spite	O
of	O
the	O
potential	O
negative	O
consequences	O
.	O
</s>
<s>
In	O
addition	O
,	O
Mostaque	O
argues	O
that	O
the	O
intention	O
behind	O
the	O
open	O
availability	O
of	O
Stable	B-General_Concept
Diffusion	I-General_Concept
is	O
to	O
end	O
corporate	O
control	O
and	O
dominance	O
over	O
such	O
technologies	O
,	O
who	O
have	O
previously	O
only	O
developed	O
closed	O
AI	O
systems	O
for	O
image	O
synthesis	O
.	O
</s>
<s>
This	O
is	O
reflected	O
by	O
the	O
fact	O
that	O
any	O
restrictions	O
Stability	B-General_Concept
AI	I-General_Concept
places	O
on	O
the	O
content	O
that	O
users	O
may	O
generate	O
can	O
easily	O
be	O
bypassed	O
due	O
to	O
the	O
availability	O
of	O
the	O
source	O
code	O
.	O
</s>
<s>
In	O
January	O
of	O
2023	O
,	O
three	O
artists	O
:	O
Sarah	O
Andersen	O
,	O
Kelly	O
McKernan	O
,	O
and	O
Karla	O
Ortiz	O
filed	O
a	O
copyright	O
infringement	O
lawsuit	O
against	O
Stability	B-General_Concept
AI	I-General_Concept
,	O
Midjourney	B-Application
,	O
and	O
DeviantArt	O
,	O
claiming	O
that	O
these	O
companies	O
have	O
infringed	O
the	O
rights	O
of	O
millions	O
of	O
artists	O
by	O
training	O
AI	O
tools	O
on	O
five	O
billion	O
images	O
scraped	O
from	O
the	O
web	O
without	O
the	O
consent	O
of	O
the	O
original	O
artists	O
.	O
</s>
<s>
The	O
same	O
month	O
,	O
Stability	B-General_Concept
AI	I-General_Concept
was	O
also	O
sued	O
by	O
Getty	O
Images	O
for	O
using	O
its	O
images	O
in	O
the	O
training	O
data	O
.	O
</s>
<s>
Unlike	O
models	O
like	O
DALL-E	B-General_Concept
,	O
Stable	B-General_Concept
Diffusion	I-General_Concept
makes	O
its	O
source	B-License
code	I-License
available	I-License
,	O
along	O
with	O
the	O
model	O
(	O
pretrained	O
weights	O
)	O
.	O
</s>
<s>
The	O
licence	O
prohibits	O
certain	O
use	O
cases	O
,	O
including	O
crime	O
,	O
libel	O
,	O
harassment	O
,	O
doxing	B-Application
,	O
"	O
exploiting	O
...	O
minors	O
"	O
,	O
giving	O
medical	O
advice	O
,	O
automatically	O
creating	O
legal	O
obligations	O
,	O
producing	O
legal	O
evidence	O
,	O
and	O
"	O
discriminating	O
against	O
or	O
harming	O
individuals	O
or	O
groups	O
based	O
on	O
...	O
social	O
behavior	O
or	O
...	O
personal	O
or	O
personality	O
characteristics	O
...	O
 [ or ] 	O
legally	O
protected	O
characteristics	O
or	O
categories	O
"	O
.	O
</s>
