Starting a news thread just in case ABX is right. ;-)

Please reply to this - if the cross-posting indeed expires i will repost 
the thread summary.

Gilles Tran wrote:
> 
> These are very good looking. On my computer the first one takes 7m50 so it's
> still pretty fast. I started the second one but stopped it after a while and
> I guess it should take somewhere between 1h and 1h30.
> The difference in render time doesn't seem much justified between both
> images as the second one still have large corner artifacts.
> 

There is indeed not much of a difference between those two at the first 
glance but the faster one, although it renders quite fast, is quite far 
away from the 'correct' solution.  If you compare the images directly 
you will see what i mean.

> However, the impressive lack of artifacts on the walls (similar to the one
> observed with C4D) leads to an obvious question : why do the default
> parameters in this case lead to so-so results? Is the trade-off the corner
> artifact problem? Could POV-Ray's radiosity be "adaptive"?

It depends on what kind of results you want.  The aim of any radiosity 
algorithm should be to converge to the correct solution if you increase 
the quality settings.  How it looks in a fast render much depends on the 
settings in POV-Ray.  To me it seems the default settings usually lead 
to a fairly uniform result, meaning the artefacts have similar intensity 
in all parts of the scene.

> Here's a C4D version obtained by boosting the "max resolution" parameter.
> This drove the render time to almost 2 hours but all the artifacts are gone
> except a couple. In fact I think I could get rid of them entirely by
> lowering the "min resolution" parameter but more tests are required. The top
> image is a screen shot of the pre-pass with the sample location visible.
> http://www.oyonale.com/temp/testrad3h.png

A similar view with POV-Ray using the previous 'e' settings:

http://www.tu-bs.de/~y0013390/files/rad_01e3.png
http://www.tu-bs.de/~y0013390/files/rad_01e2.png

Note because of the techique used to display the points the sample 
locations are not as clearly visible as in the C4D image.  If you 
extract the sample locations from the cache file and display the points 
this might lead to better results.

In rad_01e2.png the blue areas show the places where additional samples 
would be required during final trace without 'always_sample off'.  The 
small red stripes in the corners are those parts were additional samples 
were taken despite 'always_sample off'.

If you turn on always_sample you will see that the artefacts are 
slightly diminished (esp. lower right corner) but the render time is 
increased by more than 50% (28846 instead of 21738 samples)

Christoph

-- 
POV-Ray tutorials, include files, Sim-POV,
HCR-Edit and more: http://www.tu-bs.de/~y0013390/
Last updated 25 Oct. 2003 _____./\/^>_*_<^\/\.______

Post a reply to this message

news:3j4### [at] tritonimagicode...

> It depends on what kind of results you want.  The aim of any radiosity
> algorithm should be to converge to the correct solution if you increase
> the quality settings.  How it looks in a fast render much depends on the
> settings in POV-Ray.  To me it seems the default settings usually lead
> to a fairly uniform result, meaning the artefacts have similar intensity
> in all parts of the scene.

To be fair, POV-Ray default radiosity usually tend to work wonders and when
the scenes are messy enough, once can get away with very low quality values.
It's only in a few cases like this is that the radiosity problem seems
harder to solve for POV-Ray. This is why I tested C4D on this, as one of my
reasons for turning to C4D is indeed to be able to render complex interior
scenes for which I couldn't find a viable (timewise/memorywise) solution in
POV-Ray. Of course, there will be horrible trade-offs that I'm not aware of
yet in C4D but then I still have POV-Ray ;-)

> If you turn on always_sample you will see that the artefacts are
> slightly diminished (esp. lower right corner) but the render time is
> increased by more than 50% (28846 instead of 21738 samples)

What seems certain now is that both POV-Ray and C4D use the same basic
algorithm. C4D also has an "always_sample off" option (didn't try it). What
it doesn't have is the equivalent of mosaic preview (it's always like a
pixel per pixel pretrace). It does have, when the quality is high, a lag
time before the pretrace starts (POV-Ray does this too). It can do partial
radiosity renders (in the modelling view only - POV is much more flexible
when it comes to partial renderings, but I'll explain more when I know C4D
enough to make educated comparisons). I'm not sure how far these comparisons
can go anyway, because part of the speed difference could very well be
explained not by a faster radiosity algorithm, but simply by the use of scan
line.

G.

-- 

**********************
http://www.oyonale.com
**********************
- Graphic experiments
- POV-Ray and Poser computer images
- Posters

Post a reply to this message

Gilles Tran wrote:
> 
> To be fair, POV-Ray default radiosity usually tend to work wonders and when
> the scenes are messy enough, once can get away with very low quality values.
> It's only in a few cases like this is that the radiosity problem seems
> harder to solve for POV-Ray.

We should be fair and note that we probably have hundred times more 
experience to set up radiosity in POV than you have in C4D. :-)

I have tried two more renders: the first to get an idea of how the 
'correct' solution would look like: count 1600, e-b. 0.1:

http://www.tu-bs.de/~y0013390/files/rad_01z1.png

And the following uses the 'e'-Settings previously mentioned with the 
following changes:

- always_sample is on
- count is 600
- the sample direction set is randomly rotated for every sample taken 
(based on an idea by Michael Andrews)
- the count value is increased for the final render to 1200 (which 
results in the additional samples taken in the parts marked blue in

http://www.tu-bs.de/~y0013390/files/rad_01e2.png

to be gathered with this count value.  Of course this is significantly 
slower (~11min here) but the result:

http://www.tu-bs.de/~y0013390/files/rad_01g1.png

seems quite comparable to your latest C4D test concerning artefacts.

Note the C4D are much closer to the correct result than this but for a 
nice looking image this is not the most important thing.

Some additional notes: Of course these settings are especially tweaked 
for this scene - for actual use it would be nice to have settings that 
are easy to adjust without doing dozens of test renders.  The 'double 
count' trick can be simulated in standard POV using a two pass render 
and simply changing the count value in the scene file.  Of course it 
also depends on the low_error_factor technique used - i don't know the 
effect if you use it with other values.  The random rotation of samples 
was originally developed by Michael Andrews for a scene with a linear 
light source (neon tube) for which this is a very useful technique.  For 
many indoor scenes with flat walls this is probably useful although it 
also can lead to worse results in other situations.

> What seems certain now is that both POV-Ray and C4D use the same basic
> algorithm. C4D also has an "always_sample off" option (didn't try it). What
> it doesn't have is the equivalent of mosaic preview (it's always like a
> pixel per pixel pretrace). It does have, when the quality is high, a lag
> time before the pretrace starts (POV-Ray does this too). It can do partial
> radiosity renders (in the modelling view only - POV is much more flexible
> when it comes to partial renderings, but I'll explain more when I know C4D
> enough to make educated comparisons). I'm not sure how far these comparisons
> can go anyway, because part of the speed difference could very well be
> explained not by a faster radiosity algorithm, but simply by the use of scan
> line.

I am not sure if C4D uses scanline techniques for this kind of render at 
all.  And if it does it will probably not change much because more than 
90% of the time are spent for radiosity calculations and the shadows are 
probably raytraced as well.  Also note in a real scene with very 
detailed geometry the advantages of scanline are lower.  Not to mention 
reflecting and refraction objects.

Christoph

-- 
POV-Ray tutorials, include files, Sim-POV,
HCR-Edit and more: http://www.tu-bs.de/~y0013390/
Last updated 25 Oct. 2003 _____./\/^>_*_<^\/\.______

Post a reply to this message