> I've attached two images: "test1" has samples 10, "test2" has samples 200.

I would guess that what is happening is in test1, the first sample of a ray
is well lit and the other 9 are poorly lit or dark (due to self shadowing).
While in test2, the first sample is well lit and the other 199 are poorly
lit or dark. So the percentage of lit samples in the first is much greater,
causing it to be significantly brighter. This is only happening because the
media is so dense that it becomes dark at a very shallow level.

I would assume that the greater number of samples is producing the more
accurate result.

One thing to try: make two nested spheres, the inner one about 4/5 the
diameter of the outer one. Give them the same media, but give the outer one
much more samples. Then the edge of the media (in the outer sphere) will be
well sampled - as it needs to be to avoid this problem, but the inside of
the media (the inner sphere) will be poorly sampled, which is ok since it's
all dark in there anyway. If this works, it may be the fastest way to get
the effect with accuracy. (Be sure, of course, to avoid having the sphere's
volume's overlap (use a difference on the outer sphere to do this), and to
avoid coincident surfaces.)

> In my experiments, I did not see any significant difference in the output
by
> fiddling with aa. According to your explanation, "samples 10 aa_level 8
> aa_threshold 0.0001" should give similar results to "samples 160 aa_level
4
> aa_threshold 0.0001", and I don't see that with the above example. Could
> you give a more clear example?

When the media has a constant density, the first case will take 10 samples
and the latter case will take 160 samples. Only if the aa_threshold is set
to 0 (or if the media has a very noisy density) will the two samplings be
equivalent.

 - Slime
 [ http://www.slimeland.com/ ]

Post a reply to this message

Slime wrote:
> You should see the thread "default media setting inconsistancies (?)" on
> 12/27/2004 (about 15 threads down). Assuming I gave correct information in
> that thread (and I may be wrong; no one's confirmed), it should clarify
> things for you.

Its still clear as mud. Whats needed is for someone with a really good 
understanding of how media works in POV-Ray to chime in and let us all 
know whats what.

-- 
~Mike

Post a reply to this message

Among other things, Slime saw fit to write:

>> I've attached two images: "test1" has samples 10, "test2" has samples
>> 200.

(Sorry for attaching images to this non-binary group. I thought this was
p.b.i. and it was too late when I realized it wasn't)

> I would guess that what is happening is in test1, the first sample of a
> ray is well lit and the other 9 are poorly lit or dark (due to self
> shadowing). While in test2, the first sample is well lit and the other 199
> are poorly lit or dark. So the percentage of lit samples in the first is
> much greater, causing it to be significantly brighter. This is only
> happening because the media is so dense that it becomes dark at a very
> shallow level.
>
> I would assume that the greater number of samples is producing the more
> accurate result.

That was my impression and that could be the reason. There's little
difference between "samples 200" and "samples 2000"

> One thing to try: make two nested spheres, the inner one about 4/5 the
> diameter of the outer one. Give them the same media, but give the outer
> one much more samples. Then the edge of the media (in the outer sphere)
> will be well sampled - as it needs to be to avoid this problem, but the
> inside of the media (the inner sphere) will be poorly sampled, which is ok
> since it's all dark in there anyway. If this works, it may be the fastest
> way to get the effect with accuracy. (Be sure, of course, to avoid having
> the sphere's volume's overlap (use a difference on the outer sphere to do
> this), and to avoid coincident surfaces.)

Hmm... It didn't seem to be quicker in this case, the effect of the samples
was very similar and the inner surface was visible (in fact, the inner
media didn't have much influence).

I guess it's best to have as few samples as possible, as long as there are
enough to accurately sample the media (i.e. the appearance doesn't change
much with increasing samples)

> When the media has a constant density, the first case will take 10 samples
> and the latter case will take 160 samples. Only if the aa_threshold is set
> to 0 (or if the media has a very noisy density) will the two samplings be
> equivalent.

Unfortunately, povray doesn't like "aa_threshold 0". So, it seems the
adaptive sampling refers to the possible variation in the media density,
not to the effect on the image.

-- 
light_source{9+9*x,1}camera{orthographic look_at(1-y)/4angle 30location
9/4-z*4}light_source{-9*z,1}union{box{.9-z.1+x clipped_by{plane{2+y-4*x
0}}}box{z-y-.1.1+z}box{-.1.1+x}box{.1z-.1}pigment{rgb<.8.2,1>}}//Jellby

Post a reply to this message

While it is probably bad form to continue one thread in another, I'll do 
it anyway here. :)

Slime, you mentioned some method 3 default settings which you understood 
to be different from the method 1 defaults (but not listed in the docs). 
Those were intervals 1 and (min) samples 5. These values, by the way, 
are explicitly set in many of the media example scenes distributed with POV.

Anyway, I fiddled around with a couple of different settings on the 
example test scene given earlier in the thread. Going by the media 
samples listed in the render stats I get this:

method=3
intervals=default (not explicitly set)
samples=default (not explicitly set)
Media Samples: 2668281

method=3
intervals=10 (default as per the docs)
samples=1 (default as per the docs)
Media Samples: 30730560

method=3
intervals=1
samples=5
Media Samples: 1515497

method=3
intervals=1
samples=10
Media Samples: 2668281

Which leads me to conclude that the method 3 default intervals is 1 and 
samples 10, if the media samples in the render stats are an appropriate 
indicator.


-Abe

Post a reply to this message

Jellby wrote:

> 
> 
>>| But I see the opposite,
>>| as if the influence of the samples was "averaged" (since most
>>samples are
>>| deep inside, the effect of the surface ones is weaker). I would have
>>| thought that the sampling algorithm (and colour calculation) would be
>>| smarter than that, but maybe that's not so easy...
>>|
>>Do you have a sample code that produces this effect?
> 
> 
> Yep, try this (povray 3.6.1, custom compile):
> 
> ---------------------------------------
> camera { location <0, 0, -10> look_at 0 }
> 
> light_source { <-20, 20, -20>, rgb 1 }
> 
> sphere {
>   0, 3
>   hollow
>   pigment { color transmit 1 }
>   interior {
>     media {
>       scattering { 5, rgb 3 }
>       method 3
>       intervals 1
>       samples 10 // try also with 200
>     }
>   }
> }
> ---------------------------------------
> 
> I've attached two images: "test1" has samples 10, "test2" has samples 200.
> 
> 
>>It takes some trial and error, but it should be possible to get the
>>same effect with a much faster render. Of course, it all depends on
>>whether you can afford to waste more time fiddling with the
>>parameters or simply letting the computer render in the background
>>while you're doing something else ;) To make accurate comparisons, 
>>you also need to make sure that the maximum number of samples is the
>>same in both cases (I'm not talking of the second parameter of the
>>"samples" keyword here, but of the number of potential samples which
>>is something in the order of min_samples * 2^aa_level). Then if you
>>set aa_threshold to 0 the results should be the same (and take
>>*forever* to render) so all you need to do is find a value that's low
>>enough to get the effect you want while not being exactly 0.
> 
> 
> In my experiments, I did not see any significant difference in the output by
> fiddling with aa. According to your explanation, "samples 10 aa_level 8
> aa_threshold 0.0001" should give similar results to "samples 160 aa_level 4
> aa_threshold 0.0001", and I don't see that with the above example. Could
> you give a more clear example?
> 
> 
>>Note that this only applies to subsurface scattering simulations: eg
>>in cases when you have a very dense nearly uniform media.
> 
> 
> Yes, my experiments are confined to uniform media. I wonder if this adaptive
> sampling and antialiasing is only valid for non-uniform media (opposite to
> your sentence above), maybe uniform media makes adaptive sampling
> unnecessary...
> 
> Or maybe I'm making some mistakes.
> 
> 
> 
> ------------------------------------------------------------------------
> 
> 
> ------------------------------------------------------------------------
> 

Interestingly, in this particular case, method 2 with intervals 1 and 
samples 100 (or possibly less) gets pretty close to method 3 with 
samples set to 200.

Abe

Post a reply to this message

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Slime wrote:
| I would guess that what is happening is in test1, the first sample
of a ray
| is well lit and the other 9 are poorly lit or dark (due to self
shadowing).
| While in test2, the first sample is well lit and the other 199 are
poorly
| lit or dark. So the percentage of lit samples in the first is much
greater,
| causing it to be significantly brighter. This is only happening
because the
| media is so dense that it becomes dark at a very shallow level.
|
	You are essentially correct, except that the adaptive sampling
*should* take care of that. However it looks like there are some
speed optimizations in POV that prevent it from working in the case
of a constant density media (see my reply to Jellby).

		Jerome
- --
******************************
*      Jerome M. Berger      *
* mailto:jbe### [at] ifrancecom *
*  http://jeberger.free.fr/  *
******************************
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFB2xlJqIYJdJhyixIRAuamAKCO+9IqhkrMF30DUBrqLovNIW3vxACfdR5Z
QpMMYDPJj3r4xCDVCM1WBo0=
=uyYr
-----END PGP SIGNATURE-----

Post a reply to this message

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Jellby wrote:
| In my experiments, I did not see any significant difference in the
output by
| fiddling with aa. According to your explanation, "samples 10 aa_level 8
| aa_threshold 0.0001" should give similar results to "samples 160
aa_level 4
| aa_threshold 0.0001", and I don't see that with the above example.
Could
| you give a more clear example?
|
	Like Slime said, you'd need to have aa_threshold set to 0 to get the
same results (if pov accepted). You can however get the same effect
by using a ridiculously small value. Moreover I wrote a little too
fast yesterday: the precise formula is (min_samples-1) * 2^aa_level +
1, assuming of course that min_samples is at least 3.

|
|>Note that this only applies to subsurface scattering simulations: eg
|>in cases when you have a very dense nearly uniform media.
|
|
| Yes, my experiments are confined to uniform media. I wonder if this
adaptive
| sampling and antialiasing is only valid for non-uniform media
(opposite to
| your sentence above), maybe uniform media makes adaptive sampling
| unnecessary...
|
| Or maybe I'm making some mistakes.
|
	It looks like povray does some optimizations in the case of a
constant density media. I did a quick check by adding the following
density to your code:

density {
~  spherical
~  color_map {
~    [ 0.0 rgb 1.0 ]
~    [ 1.0 rgb 1.0 ]
~  }
}

	Since the colormap is rgb 1.0 everywhere, this should give the same
results as the original but it doesn't. The result is pretty much
what I described in my previous post (except that samples 129 takes
much longer to render since there are a lot of pure black samples
taken that get culled in the samples 3 case). You may take a look for
example at those pictures done with aa_threshold 1e-200:

aa_level 8
samples 3
http://jeberger.free.fr/tmp/media-3-8.png
aa_level 2
samples 129
http://jeberger.free.fr/tmp/media-129-2.png

	I'd say that this is a bug.

	Note that you can probably get the correct result with a faster
render time if you disable this optimization by adding a pseudo
density like I just did...

		Jerome

PS: If the pictures aren't there, try again a bit later: as of this
posting, the second one hadn't finished rendering...

- --
******************************
*      Jerome M. Berger      *
* mailto:jbe### [at] ifrancecom *
*  http://jeberger.free.fr/  *
******************************
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFB2xn0qIYJdJhyixIRAq1IAJ0aa4m+mH6jtx8gp/ak293dlMu1fACgnglN
9sSSzvW20+Za6LOJn1rGj34=
=/Pz9
-----END PGP SIGNATURE-----

Post a reply to this message

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

	If your file doesn't contain a #version directive, or if the version
specified is higher than 3.5, then the default for media are:
method 3
intervals 1
samples 10, 10

	Note that this means that if you only change the method, you'll
still get 1 interval and 10 samples unless you change them too. For
compatibility reason, if your file has a #version lower than 3.5, the
defaults are those written in the doc. I guess that the doc should be
updated...

		Jerome
- --
******************************
*      Jerome M. Berger      *
* mailto:jbe### [at] ifrancecom *
*  http://jeberger.free.fr/  *
******************************
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFB2x2FqIYJdJhyixIRAmM+AJsFL9V6ng9mxywUUXIbNEiTTNEg5ACgoWT3
xJak+RLSsidTwNhWOKkDikc=
=sZfU
-----END PGP SIGNATURE-----

Post a reply to this message

> You are essentially correct, except that the adaptive sampling
> *should* take care of that.

I was under the impression that the adaptive sampling only took into account
the values of the media density, and not the lighting.

 - Slime
 [ http://www.slimeland.com/ ]

Post a reply to this message

> Which leads me to conclude that the method 3 default intervals is 1 and
> samples 10, if the media samples in the render stats are an appropriate
> indicator.

Ah, thanks.

 - Slime
 [ http://www.slimeland.com/ ]

Post a reply to this message