Content-Transfer-Encoding: quoted-printable
Content-Type: text/plain; charset="utf-8"
User-Agent: Workspace Webmail 5.6.46
Message-Id: <20140305112731.89259f98f291d771214680e8971719cb.e67b5c1e9b.wbe@email04.secureserver.net>
From: "Robert W. Walker" <robertw@walkersound.com>
To: "After Effects Mail List" <AE-List@media-motion.tv>
Subject: RE: [AE] Using optical flow to speed up 3D renders
Date: Wed, 05 Mar 2014 11:27:31 -0700
Mime-Version: 1.0

Two weeks ago I had a major deadline problem that Twixtor helped me=0Asolve=
, with a caveat.=0A=0AI was rendering sequences of photo-realistic CGI prod=
uct shots. Global=0Aillumination settings high enough to make the renders c=
lean and gorgeous=0Aalso incurred heinous render times. Doing the per-frame=
 render math, I=0Agulped. I was going to finish rendering a day and a half =
after delivery=0Awas due. I needed to cut the render time in half, so I did=
 a short test=0Arender of a couple of seconds at TWELVE fps, applied Twixto=
r and it=0Aworked.=0A=0AThe caveat: I got severe artifacting on frames that=
 had fast-moving=0Aforeground elements against relatively static mid-ground=
 elements.=0AFortunately the segments of these particular scenes included o=
nly a=0Ahandful frames where that happened so I rendered those out at 24fps=
 and=0Awas able to make it work by simply cutting those frames into the=0AT=
wixtor'ed 12fps footage. It was so quick you couldn't tell.  Prolonged=0Afo=
reground/background motion differential would be a bigger problem,=0Athough=
 from what I could see.=0A=0ARobert W. Walker=0ALos Angeles=0A=0A=0A> -----=
--- Original Message --------=0A> Subject: Re: [AE] Using optical flow to s=
peed up 3D renders=0A> From: Stephen van Vuuren <stephen@sv2studios.com>=0A=
> Date: Wed, March 05, 2014 12:34 pm=0A> To: "After Effects Mail List" <AE-=
List@media-motion.tv>=0A> =0A> =0A> I have not used optical flow for speedi=
ng up 3D renders specifically but I've used Twixtor for a number of project=
s over the years with various types of sources that have included some 3D s=
ource material.=0A> =0A> The problem with optical flow is that it's not an =
automatic solution i.e. set and forget as camera motion, foreground and bac=
kground layers can confuse and cause subtle or obvious artifacts. =0A> =0A>=
 On say a two second shot with a lot of camera and subject movement (especi=
ally if CGI has handheld or zoom camera simulation), it's likely manual twe=
aking of the optical flow is going to be necessary even with tweaked motion=
 vectors.=0A> =0A> If the 3D CGI renders are fairly static and predictable =
(for the optical flow) it's possible to get settings that will work for a w=
hole shot. =0A> =0A> But the reason optical flow is simply not a default so=
lution is that ultimately computers and software is still stupid and can't =
compensate for unexpected situations.=0A> =0A> And while it's much faster t=
han it used to be, unless the CGI renders are brutal, the renders are still=
 going to be considerable with optical flow required to get artifact free p=
rocessing. Plus manual labor time.=0A> =0A> Again - certain controlled scen=
arios it might work - but unlikely to be a simple plug and play solution.=
=0A> =0A> stephen van vuuren=0A> 336.202.4777=0A> =0A> http://www.insaturns=
rings.com/=0A> http://www.sv2dcp.com/=0A> http://www.sv2studios.com/=0A> =
=0A> A film is - or should be - more like music than like fiction. It shoul=
d be a progression of moods and feelings. The theme, what's behind the emot=
ion, the meaning, all that comes later.=0A> -Stanley Kubrick=0A> =0A> -----=
Original Message-----=0A> From: After Effects Mail List [mailto:AE-List@med=
ia-motion.tv] On Behalf Of Chris Zwar=0A> Sent: Wednesday, March 05, 2014 6=
:48 AM=0A> To: After Effects Mail List=0A> Subject: [AE] Using optical flow=
 to speed up 3D renders=0A> =0A> I have this vague recollection that optica=
l flow technology (e.g. twixtor and reel smart motion blur) was originally =
developed to help speed up 3D rendering workflows.  I think the idea was th=
at 3D animators could render out every 2nd or 3rd frame, and use an optical=
 flow plugin like twixtor to create the in-betweens.  Even if that's not st=
rictly correct, the potential is there.  Some of the photo-realistic projec=
ts I've worked on have had 3D render times of 3 - 5 hours a frame.  Renderi=
ng every 2nd frame is effectively halving the overall render time, which ca=
n be a massive saving.  Even a slow After Effects plugin is usually only se=
conds per frame, not hours.=0A> =0A> So I was wondering if anyone has actua=
lly done this, or tried using other 2D techniques to help speed up 3D rende=
ring.  =0A> =0A> I can think of 3 ways in which slow 3D renders can be comp=
ensated for by faster compositing techniques:=0A> =0A> 1) Up-resing.  For e=
xample rendering at 720p instead of 1080p and scaling up the finished rende=
rs.  If compositing multiple passes, only the slow renders need to be small=
er and scaled up.=0A> 2) De-noising.   Forgive me for not knowing the corre=
ct terminology, but when rendering with global illumination it seems that t=
here's an overall quality setting that directly determines both the speed o=
f rendering and the noisiness of the image.  Rendering with a lower setting=
 can make renders noisier, but a de-noising plugin such as Neat Video can f=
ix this.=0A> 3) As stated above, rendering every 2nd or 3rd frame and using=
 something like twixtor to create the missing frames.  A motion vector pass=
 would make this more accurate.=0A> =0A> So I'm familiar with 2 of those 3 =
approaches - I have worked in situations where 3D passes are rendered at sm=
aller sizes and then scaled up.  It works very well and the time savings ca=
n be dramatic when dealing with very long renders.  3D renders can be so cl=
ean that they scale up very well.=0A> =0A> I have also worked in situations=
 where the neat video de-noiser was used to compensate for noisy GI renders=
, and again the savings can be dramatic - in some cases this can almost hal=
f 3D rendering times.  Neat video seems to be an incredible plugin, so much=
 faster and so much better than the AE equivalent.=0A> =0A> So that leaves =
the optical flow technique as the one I haven't tried yet.  Has anyone done=
 this?  I'd love to hear from real-world examples where people were able to=
 render every 2nd or 3rd frame.  Is a motion vector pass essential for it t=
o work properly?=0A> =0A> Any other thoughts or insight welcome...=0A> =0A>=
 -Chris=0A> +---End of message---+=0A> To unsubscribe send any message to <=
ae-list-off@media-motion.tv>=0A> =0A> +---End of message---+=0A> To unsubsc=
ribe send any message to <ae-list-off@media-motion.tv>