<s>
In	O
machine	O
learning	O
,	O
diffusion	B-Algorithm
models	I-Algorithm
,	O
also	O
known	O
as	O
diffusion	B-Algorithm
probabilistic	I-Algorithm
models	I-Algorithm
,	O
are	O
a	O
class	O
of	O
latent	O
variable	O
models	O
.	O
</s>
<s>
The	O
goal	O
of	O
diffusion	B-Algorithm
models	I-Algorithm
is	O
to	O
learn	O
the	O
latent	O
structure	O
of	O
a	O
dataset	O
by	O
modeling	O
the	O
way	O
in	O
which	O
data	O
points	O
diffuse	O
through	O
the	O
latent	B-Algorithm
space	I-Algorithm
.	O
</s>
<s>
In	O
computer	B-Application
vision	I-Application
,	O
this	O
means	O
that	O
a	O
neural	O
network	O
is	O
trained	O
to	O
denoise	O
images	O
blurred	O
with	O
Gaussian	O
noise	O
by	O
learning	O
to	O
reverse	O
the	O
diffusion	O
process	O
.	O
</s>
<s>
Three	O
examples	O
of	O
generic	O
diffusion	O
modeling	O
frameworks	O
used	O
in	O
computer	B-Application
vision	I-Application
are	O
denoising	O
diffusion	B-Algorithm
probabilistic	I-Algorithm
models	I-Algorithm
,	O
noise	O
conditioned	O
score	O
networks	O
,	O
and	O
stochastic	O
differential	O
equations	O
.	O
</s>
<s>
Diffusion	B-Algorithm
models	I-Algorithm
were	O
introduced	O
in	O
2015	O
with	O
a	O
motivation	O
from	O
non-equilibrium	O
thermodynamics	O
.	O
</s>
<s>
Diffusion	B-Algorithm
models	I-Algorithm
can	O
be	O
applied	O
to	O
a	O
variety	O
of	O
tasks	O
,	O
including	O
image	O
denoising	O
,	O
inpainting	B-Algorithm
,	O
super-resolution	B-Algorithm
,	O
and	O
image	B-General_Concept
generation	I-General_Concept
.	O
</s>
<s>
For	O
example	O
,	O
an	O
image	B-General_Concept
generation	I-General_Concept
model	O
would	O
start	O
with	O
a	O
random	O
noise	O
image	O
and	O
then	O
,	O
after	O
having	O
been	O
trained	O
reversing	O
the	O
diffusion	O
process	O
on	O
natural	O
images	O
,	O
the	O
model	O
would	O
be	O
able	O
to	O
generate	O
new	O
natural	O
images	O
.	O
</s>
<s>
Announced	O
on	O
13	O
April	O
2022	O
,	O
OpenAI	O
's	O
text-to-image	B-General_Concept
model	I-General_Concept
DALL-E	B-General_Concept
2	I-General_Concept
is	O
a	O
recent	O
example	O
.	O
</s>
<s>
It	O
uses	O
diffusion	B-Algorithm
models	I-Algorithm
for	O
both	O
the	O
model	O
's	O
prior	O
(	O
which	O
produces	O
an	O
image	O
embedding	O
given	O
a	O
text	O
caption	O
)	O
and	O
the	O
decoder	O
that	O
generates	O
the	O
final	O
image	O
.	O
</s>
<s>
Consider	O
the	O
problem	O
of	O
image	B-General_Concept
generation	I-General_Concept
.	O
</s>
<s>
As	O
it	O
turns	O
out	O
,	O
allows	O
us	O
to	O
sample	O
from	O
using	O
stochastic	B-Algorithm
gradient	I-Algorithm
Langevin	I-Algorithm
dynamics	I-Algorithm
,	O
which	O
is	O
essentially	O
an	O
infinitesimal	O
version	O
of	O
Markov	B-General_Concept
chain	I-General_Concept
Monte	I-General_Concept
Carlo	I-General_Concept
.	O
</s>
<s>
Taking	O
the	O
perspective	O
of	O
the	O
noisy	B-General_Concept
channel	I-General_Concept
model	I-General_Concept
,	O
we	O
can	O
understand	O
the	O
process	O
as	O
follows	O
:	O
To	O
generate	O
an	O
image	O
conditional	O
on	O
description	O
,	O
we	O
imagine	O
that	O
the	O
requester	O
really	O
had	O
in	O
mind	O
an	O
image	O
,	O
but	O
the	O
image	O
is	O
passed	O
through	O
a	O
noisy	O
channel	O
and	O
came	O
out	O
garbled	O
,	O
as	O
.	O
</s>
<s>
Image	B-General_Concept
generation	I-General_Concept
is	O
then	O
nothing	O
but	O
inferring	O
which	O
the	O
requester	O
had	O
in	O
mind	O
.	O
</s>
<s>
In	O
other	O
words	O
,	O
conditional	O
image	B-General_Concept
generation	I-General_Concept
is	O
simply	O
"	O
translating	O
from	O
a	O
textual	O
language	O
into	O
a	O
pictorial	O
language	O
"	O
.	O
</s>
<s>
The	O
classifier-guided	O
diffusion	B-Algorithm
model	I-Algorithm
samples	O
from	O
,	O
which	O
is	O
concentrated	O
around	O
the	O
maximum	B-General_Concept
a	I-General_Concept
posteriori	I-General_Concept
estimate	I-General_Concept
.	O
</s>
<s>
In	O
the	O
context	O
of	O
diffusion	B-Algorithm
models	I-Algorithm
,	O
it	O
is	O
usually	O
called	O
the	O
guidance	O
scale	O
.	O
</s>
<s>
This	O
is	O
an	O
integral	O
part	O
of	O
systems	O
like	O
GLIDE	O
,	O
DALL-E	B-General_Concept
and	O
Google	O
Imagen	O
.	O
</s>
