Warp wrote:
>   It would be nice to see some rendertimes.

I have been simply amazed by these "new breed of" renderers which use
unbiased methods that make the image converge slowly to the correct
solution. Just look at these amazing images made with indigo Renderer:

http://www.indigorenderer.com/joomla/index.php?option=com_gallery2&Itemid=26

The rendering times usually vary from few hours to tens of hours. But it
all depends on quality (ie. noise) requirements. You get a preview very
quickly with a lot of noise.

I've been implementing my own path tracer and my impressions are:

I love the fact that you always get (almost) all the characteristics of
light taken into account. You don't have to figure whether radiosity,
caustics, indirect shadows, indirect refractions etc.etc. play
significant role or not. You don't have to tweak tens of parameters and
guessimate which parameters/values/features give you the effect that the
renderer should handle in the first place. And you don't have to guess
if the artifact you see is really an artifact (because wrong parameters
or whatever) or correct result.

I just implemented specular reflections.After doing those I noticed:
"Wow, I just accidentally implemented caustics and phong shading". The
same way you don't have to guess which kind of light source and which
parameters you must use to get the desired effect. Every lightsource can
be of any shape (not just 2d fake sources or physically impossible
points) and every point on the light source surface does indeed emit
light. No parameters, no features to enable, no shadow artifacts. It
Just Works (tm).

The whole purpose of raytracer is to give you an accurate image.
Otherwise you could use faster scanline renderers. Brute-force renderers
really accomplish this as can be seen from the images. IMO (and this is
just my opinion) you quite rarely see images of such quality made with
Pov - at least compared to the ones rendered with Indigo. (By "quality"
I mean lighting realism - not necessarily model quality etc). In many
cases this is not because POV wasn't capable of producing almost as good
images. The reason is that it is difficult and time consuming to
guess/try/tweak which features/parameters you have to enable/set to get
realistic results. Many images made by new users look worse than 3D
games that are 5 years old. Had they used brute-force, the results would
be a lot more photorealistic with little effort.

The rendering time is not a that big problem in still images -
animations are a different matter. Brute-force renderers are very easily
made multi-threaded (I know because I did it just 3 hours ago...) and
thus they scale very well with new multi core CPUs.

I waited 3 hours for a small refractive sphere to get drawn on my screen
back in the 90s. I can wait the same 3 hours now to get true global
illumination, caustics and other very important effects of light and I
can be 100% certain that the result is accurate.

I also hope these new methods are considered when rewriting/designing of
Pov4 begins.

This was not a rant against POV. But things go forward and methods that
were impossible to use 10-20 years ago are now very usable and superior
in many regards. And after 5 years they are even more usable.

SS

Post a reply to this message

Severi Salminen <sev### [at] notthissaunalahtifiinvalid> wrote:
> I waited 3 hours for a small refractive sphere to get drawn on my screen
> back in the 90s. I can wait the same 3 hours now to get true global
> illumination, caustics and other very important effects of light and I
> can be 100% certain that the result is accurate.

  Am I incorrect if I get the impression that these unbiased renderers
offer nothing else than unbiased rendering? In other words, even the
simplest of scenes will take hours to look ungrainy, no matter what
you do?

  Sometimes I use POV-Ray to get simple 3D-looking graphics for diverse
things. The big advantage is that I can do it easily and POV-Ray renders
it very fast. We are talking about a few seconds even for large-sized
images. It would be completely counter-productive to have to wait for
hours for a simple image to look non-grainy, when all you want is something
quick which looks "cool" and "3D'ish".

  For this reason even if POV-Ray in the future supports unbiased rendering,
it should always be an *alternative* method of rendering, not the only
available one. Removing the current phongshading-based rendering would be
a setback in many areas. After all, POV-Ray is not *always* used for
physical simulations of reality.

-- 
                                                          - Warp

Post a reply to this message

yes, I saw luxrender2.org was having issues.

hmm, trying to lure users away from povray or trying to get povray 
developers hard-pressed to integrate unbiased methods into the renderer? ;)

You'll only perhaps succeed with the first goal if you somehow manage to 
get luxrender to parse povray SDL files.  Come on, RIB and all the 
current XML crap sucks very much compared to the high-level povray SDL...

I'll keep watching lux, though...

Post a reply to this message

delle wrote:
> Warp <war### [at] tagpovrayorg> wrote:
>> It would be nice to see some rendertimes.
>>
>> --
>>                                                           - Warp
> 
> I'm looking for the best Quality possible, LuxRender seems to be on the right
> path... CPU are becoming faster and faster... so...
> 
> Even Povray with high quality settings on in not so fast....
> 
> try to enable: Radiosity (hi Q), area Lights, anti aliasing and camera depth of
> field...
> 
> Speaking about "low quality", my Nvidia card can play Crysis at 30 fps ... the
> quality is quite good...

You didn't answer his question. I'm also curious about render times, but 
you completely dodged the question.

Post a reply to this message

Mike Raiford wrote:

> You didn't answer his question. I'm also curious about render times, but
> you completely dodged the question.

See the indigo site gallery. Some captions include render time. Of
course that tells you nothing about how good/bad the image looked in
less time. Remember that with brute force you see a lot even after 1 minute.

Post a reply to this message

Severi Salminen wrote:

> See the indigo site gallery. Some captions include render time. Of
> course that tells you nothing about how good/bad the image looked in
> less time. Remember that with brute force you see a lot even after 1 minute.

I looked. I remember that gallery. 18 hours on one of the archetecture. 
Essentially set it, then come back the next day, and you might have a 
nice image. :)


I dunno, These types of renderers do have some interesting applications, 
I could see an architect using them to do a true physical light 
simulation on a building design, using various lighting schemes/time of 
day, etc.. But for creating "photorealistic" rendered images, my bets 
are still on raytracers, they're not a physically accurate simulation, 
but they do a decent approximation.

Post a reply to this message

Severi Salminen a écrit :
> Mike Raiford wrote:
> 
>> You didn't answer his question. I'm also curious about render times, but
>> you completely dodged the question.
> 
> See the indigo site gallery. Some captions include render time. Of
> course that tells you nothing about how good/bad the image looked in
> less time. Remember that with brute force you see a lot even after 1 minute.

Maybe in the case of these renderers some other indicator of rendering 
time should be used. Technically you can let them run forever, the image 
is never finished, and getting better all the time...

The curve of "noisiness" vs. rendering time would be interesting. For 
some definition of "noisiness" :-)

I'm attracted by the idea of these unbiased renderers because of the 
ease of use described in another post. What holds me back is indeed the 
scene description language. And laziness ;-)

-- 
Vincent

Post a reply to this message

Warp wrote:

> Am I incorrect if I get the impression that these unbiased renderers
> offer nothing else than unbiased rendering? 

"Nothing else than perfect simulation of light". Nope, nothing else ;-)

> In other words, even the simplest of scenes will take hours to look
> ungrainy, no matter what you do?

Define "ungrainy". But yes, it is a tradeoff. You get grain, but you get
the same results no matter what features there are in the scene.
Caustics don't cost extra. Global illumination doesn't cost extra. Etc.
I can make a similar question:

With PovRay you don't see even one single renderd line with "perfect
radiosity/photon mapping/area light size, anti aliasing etc. It is a
trade-off, again.

Also, there are many ways to speed up brute force renderes. And there
are many ways to do it. Like path tracing, bidirectional path tracing
etc. You can implement various methods that reduces the amount of rays
needed for certain variance. I have yet to research those more.

>   Sometimes I use POV-Ray to get simple 3D-looking graphics for diverse
> things. The big advantage is that I can do it easily and POV-Ray renders
> it very fast. We are talking about a few seconds even for large-sized
> images. It would be completely counter-productive to have to wait for
> hours for a simple image to look non-grainy, when all you want is something
> quick which looks "cool" and "3D'ish".

I do the same thing now with my own program. I get a grainy preview
pretty quickly. And the grainy preview show perfect features. If you
simply want "3D-looking graphics" there are many a lot faster and better
options than PovRay. You discard the realism but you get it fast. I'm
more interested in photorealism.

>   For this reason even if POV-Ray in the future supports unbiased rendering,
> it should always be an *alternative* method of rendering, not the only
> available one. Removing the current phongshading-based rendering would be
> a setback in many areas. After all, POV-Ray is not *always* used for
> physical simulations of reality.

I agree that there are situations where you need speed over accuracy.
That is true. But also the opposite is true. After all, POV was created
to create realistic images - not the give decent results as fast as
possible. I'm not sure what is the general opinion: do users want to get
the best possible results in longer time - or okayish results in less time?

I belong to the first group.

I attached a very simple scene that took 10minutes to render to this
noise level. You can see caustics where reflective sphere and ground
meets. This is a simple scene where POV would be a lot faster (if you
know how to set it up properly..) Im also sure that good brute force
renderers give a lot better results in shorter time than my immature
renderer. The second image was done in 50 seconds. Just to show the
difference.

Post a reply to this message

Attachments:
Download 'kuva5.jpg' (20 KB) Download 'kuva6.jpg' (45 KB)

Preview of image 'kuva5.jpg'

Preview of image 'kuva6.jpg'

Mike Raiford wrote:

> I looked. I remember that gallery. 18 hours on one of the archetecture.
> Essentially set it, then come back the next day, and you might have a
> nice image. :)

Actually, you'll have a good image in less than 1 hour. After 18h it is
very good :=)

> I could see an architect using them to do a true physical light
> simulation on a building design, using various lighting schemes/time of
> day, etc.. But for creating "photorealistic" rendered images, my bets
> are still on raytracers, they're not a physically accurate simulation,
> but they do a decent approximation.

What do you mean "your bets are on"? Brute forcers give always better
looking results. Nobody can claim anything else. The only problem is the
time it takes to get to decent noise levels. So other than speed my bet
on photorealistic images are, of course, on brute-forcers. And we know
how the situation looks after few years of multi-core CPU development :)

Post a reply to this message

Severi Salminen <sev### [at] notthissaunalahtifiinvalid> wrote:
> I do the same thing now with my own program. I get a grainy preview
> pretty quickly. And the grainy preview show perfect features. If you
> simply want "3D-looking graphics" there are many a lot faster and better
> options than PovRay.

  How many of them are for Linux? Of those, how many of them work without
OpenGL?

  And if something takes 5 second to render instead of 1, big deal.

> You discard the realism but you get it fast. I'm
> more interested in photorealism.

  A 3D button for a webpage doesn't need any more photorealism than what
POV-Ray can offer in a 1 second render.

-- 
                                                          - Warp

Post a reply to this message