<s>
Neural	B-Algorithm
style	I-Algorithm
transfer	I-Algorithm
(	O
NST	O
)	O
refers	O
to	O
a	O
class	O
of	O
software	O
algorithms	O
that	O
manipulate	O
digital	O
images	O
,	O
or	O
videos	O
,	O
in	O
order	O
to	O
adopt	O
the	O
appearance	O
or	O
visual	O
style	O
of	O
another	O
image	O
.	O
</s>
<s>
Several	O
notable	O
mobile	O
apps	O
use	O
NST	O
techniques	O
for	O
this	O
purpose	O
,	O
including	O
DeepArt	B-Application
and	O
Prisma	B-Application
.	O
</s>
<s>
The	O
first	O
two	O
example-based	O
style	B-Algorithm
transfer	I-Algorithm
algorithms	O
were	O
image	B-Algorithm
analogies	I-Algorithm
and	O
image	O
quilting	O
.	O
</s>
<s>
The	O
original	O
paper	O
used	O
a	O
VGG-19	O
architecture	O
that	O
has	O
been	O
pre-trained	O
to	O
perform	O
object	O
recognition	O
using	O
the	O
ImageNet	B-General_Concept
dataset	O
.	O
</s>
<s>
In	O
2017	O
,	O
Google	B-Application
AI	I-Application
introduced	O
a	O
method	O
that	O
allows	O
a	O
single	O
deep	O
convolutional	O
style	B-Algorithm
transfer	I-Algorithm
network	O
to	O
learn	O
multiple	O
styles	O
at	O
the	O
same	O
time	O
.	O
</s>
<s>
These	O
activations	O
are	O
encoded	O
into	O
a	O
Gramian	B-Algorithm
matrix	I-Algorithm
representation	O
,	O
call	O
it	O
to	O
denote	O
the	O
'	O
style	O
 '	O
of	O
.	O
</s>
<s>
An	O
iterative	O
optimization	O
(	O
usually	O
gradient	B-Algorithm
descent	I-Algorithm
)	O
then	O
gradually	O
updates	O
to	O
minimize	O
the	O
loss	O
function	O
error	O
:	O
</s>
<s>
Then	O
we	O
successively	O
backpropagate	B-Algorithm
this	O
loss	O
through	O
the	O
network	O
with	O
the	O
CNN	O
weights	O
fixed	O
in	O
order	O
to	O
update	O
the	O
pixels	O
of	O
.	O
</s>
<s>
Algorithms	O
are	O
typically	O
implemented	O
for	O
GPUs	B-Architecture
,	O
so	O
that	O
training	O
takes	O
a	O
few	O
minutes	O
.	O
</s>
<s>
they	O
explored	O
the	O
fusion	O
of	O
optical	O
flow	O
information	O
into	O
feedforward	B-Algorithm
networks	I-Algorithm
in	O
order	O
to	O
improve	O
the	O
temporal	O
coherence	O
of	O
the	O
output	O
.	O
</s>
<s>
Most	O
recently	O
,	O
feature	B-Algorithm
transform	I-Algorithm
based	O
NST	O
methods	O
have	O
been	O
explored	O
for	O
fast	O
stylization	O
that	O
are	O
not	O
coupled	O
to	O
single	O
specific	O
style	O
and	O
enable	O
user-controllable	O
blending	O
of	O
styles	O
,	O
for	O
example	O
the	O
whitening	O
and	O
coloring	O
transform	O
(	O
WCT	O
)	O
.	O
</s>
