|
"filtering doesn't decide how many samples are taken..."
Hey Steven, that's exactly what I've said, reread my mails. However, the
filter size *does* decide on how many samples will be combined for a
single pixel *after* all rays are shot by creating an ellipse around the
center of that pixel and collecting all the samples within the given x/y
radii.
Your observations regarding the eye-rays are interesting, though. We
seem to have a few extra rays in there, but I see the basic idea of
samples = pixels with a sampling value of 0 still being valid, don't you
think? At least I can live with a variance of less than ten percent... ;)
Steven Caron wrote:
/deep breath!
/filtering doesn't decide how many samples are taken...
file:///C:/Softimage/XSI_5.0/Doc/mental_ray/manual/node76.html#INDEX188
"...The *filter* statement specifies how multiple samples are to be
combined into a single pixel value..."
and if you just take look at jitter thats only two definitions down
from filter...
"...Without *jittering*, samples are taken at the corners of pixels or
subpixels..."
and if you turn jitter on. with AA min 0 and max 0... mental ray
kindly turns jitter off :) why? Because it will introduce artifacts
with that low of a sampling rate...
//INFO : RC 0.2 warn 082008: jittering disabled because max
samples < 1
So according to you a 100 x 100 render would produce 10000 samples? (
eye rays ) Lets test this!
I setup a format of... 100 x 100. AA min of 0 and max of 0. Jitter
OFF. Threshold rgba to 1.0 and filter type box 1 1, but threshold
doesn't matter because of min and max samples being the same and
filter doesn't matter because it only averages samples already taken.
turn progress on...
//INFO : RC 0.2 info : rendering statistics
//INFO : RC 0.2 info : type number per
eye ray
//INFO : RC 0.2 info : eye rays
10816 1.00
hmmm interesting!
I am going to throw another variable in here... task size or tile
size. this option will really increase or decrease your samples...
smaller task size is going to give you more samples.. larger is going
to give you less. Here is a quick formula for predicting samples. I
say predicting, because there is no way to know everything they are
doing internally, short of being a mental images developer.
task size + 1 * task size + 1 = samples per tile...
To test this make a 32 x 32 resolution AA min 0 max 0 and turn your
"tile size" to 32x32. You can leave automatic in cause most likely
mental ray will choose this, but just in case force it. you know so
you feel like your in control. You will get...
32 + 1 * 32 + 1 = 1089
have fun.. do some more tests...
André Adam wrote:
To add to this, the position sampled is only important for the
filtering algorithm that collects samples using a specific radius
around the center of the pixel. With an aa setting of 0 and a box
filter with a size of 1, indeed four samples will be considered to
retrieve the final colour value for this pixel. However, hardly
anyone uses a box 1/1 filter. When switching to the more popular
Mitchell filtering, which in most setups needs higher radii to create
nice-looking imagary, many more samples are taken into consideration
for a single pixel's colour value. However, the sample density shot
by the renderer *does not* increase in this scenario, with aa 0 it
stays one sample shot per pixel.
André Adam wrote:
Nope, with aa 0 you get exactly as many samples as you have pixels
for the whole image. It definately is one sample per pixel. Not
centered, but that doesn't make much of a difference, the spacing
between the samples stays one pixel regardless of the position
sampled. It's just about offsetting the sampling raster. Which btw
gets randomised anyways when enabling the sample jitter option.
Cheers!
-André
Steven Caron wrote:
"...and one sample per pixel (aa setting 0)..."
just like bernard and kim said about threshold being misunderstood
this too is misunderstood.
mental ray samples at the corners of a pixel. meaning 4 per pixel
not 1. but not always 4 new samples because mental ray can share
samples with bordering pixels.
*bracing myself now*
steven
On 6/2/06, *André Adam* <a_adam(at)49games.de
<mailto:a_adam(at)49games.de>> wrote:
A really nice example covering this problem is a hemisphere mapped
with
a clear night sky showing lots of small stars. If the min aa
setting is
too broad (and one sample per pixel (aa setting 0) might very
well be
too broad), the rays will likely completely miss out certain
stars on
certain frames which leads to the stars wildly flickering in
animation...
kim aldis wrote:
>I'm not trying to put you down here Bernard but there are
situations where
>settings like this aren't good and you should have some idea of
what sort of
>image you're rendering before making decisions about aa settings.
In some
>shots, messing with threshold won't make any difference at all,
raising max
>settings will make little difference unless you bring up the min
settings.
>You're right, threshold is largely misunderstood. Let's try and
shed some
>light:-
>
>here's something you can try. Get a cube, unit 1. scale it so
it's really
>long and really thin - 0.02, 0.02, 35.
>now duplicate it 20 or 30 times and space each cube one unit
apart in x. Now
>shift them all back away from the camera so you get a good
convergance. Lots
>of thin, parallel lines in the distance, all converging to the
vanishing
>point is what you're looking for.
>
>Now mess with the aa settings and pay close attention to the
quality where
>the cubes are nearly converging. Try 0,3 and mess with the
threshold. Not a
>lot of difference, is there, even if you set the threshold really
as low as
>you can get it. In fact, changing the threshold varies not a lot
at all.
>
>now compare 0,2 with 0,3. Not much better. And I'll bet if you
try 1,2 it'll
>be way better than 0,3. Again, regardless of the threshold.
>
>The point here is that there's a real danger of getting it wrong
if you
>start generalising aa settings and don't take into account the
kind of image
>you're rendering, as Helen is finding out, since sampling grain
is fine
>detail. I just tried something here. Pushing aliaising to 1,2
with default
>threshold gave me better improvement than doubling the
sampling size.
>lowering the threshold improved things a bit but nothing worked
better - nor
>was faster - than raising the min setting, bringing it closer to
the max.
>
>If you really want to see how well - or otherwise - aa is
working, turn on
>'view sampling' in the region diagnostic tab. It's most
revealing.
>
>-----Original Message-----
>From: owner-xsi(at)Softimage.COM <mailto:owner-xsi(at)Softimage.COM>
[mailto:owner-xsi(at)Softimage.COM <mailto:owner-xsi(at)Softimage.COM>]
On Behalf Of
>Bernard Lebel
>Sent: 01 June 2006 16:23
>To: XSI(at)Softimage.COM <mailto:XSI(at)Softimage.COM>
>Subject: Re: shadow flicker
>
>I would definitely use a min of 0 and a max of 3.
>
>Also, what is generally ignored or misunderstood is the treshold.
The
>treshold color is a control for how sampling is done. Lower
values means
>that sampling occurs with lower contrasts, in other words, the
level of
>sampling have more chances of getting closer to the max value.
Higher
>treshold means a higher contrasts is tolerated before requiring
another
>level of sampling, so less samples may be takne.
>
>I never work with treshold higher than 0.05, and had to go as
low as
>0.005 on few occasions. Note that at that point, if the image is
detailed,
>you should consider ignoring adaptive sampling altogether and use
same min
>and max values. If the image has significant amounts of constant
colors
>(like shadow passes, black backgrounds in the likes), then
adapative
>sampling is still a viable solution.
>
>
>Lastly, my rule of thumb for the filtering is Gaussian 3/3 for TV
output and
>Mitchell 4/4 for film output. Gaussian gives a blurrier sample
>interpolation, which may "conceal" minor imperfections.
>
>
>
>Cheers
>Bernard
>
>
>
>
>On 6/1/06, Helen Bucknall <helen(at)helenbucknall.com
<mailto:helen(at)helenbucknall.com>> wrote:
>
>
>>Mostly I just use min -1 and max 2. and box filtering 1 or
sometimes I
>>use triangle 2 for a softer effect.
>>
>>I output a string of 852 x 480 bmps to my clients and then
they mpeg
>>it somehow and wack it up on big plasma screens. They never
complain.
>>When they want stills for print I just render a frame out huge
>>(3000+)but leave the antialiasing the same.
>>If it's for telly I do HD res. If it's to be comped with live
action I
>>ask what anti aliasing they want.
>>
>>
>>
>>Bernard Lebel wrote:
>>
>>
>>
>>>Samples 20 is a bit low. I generally put 64 as soon as I set
up the
>>>shadow map.
>>>
>>>What do you mean with "standard antialiasing"?
>>>
>>>
>>>Bernard
>>>
>>>
>>>
>>>
>>>On 5/31/06, Helen Bucknall <helen(at)helenbucknall.com
<mailto:helen(at)helenbucknall.com>> wrote:
>>>
>>>
>>>
>>>>OK. I dragged the spotlight in closer to the scene elements
and it
>>>>seems to have stopped all the flickering! It wasn't that far
away,
>>>>but now it is almost close enough to make my character
sweaty. I
>>>>also scaled down my ground plane a tad.
>>>>
>>>>Just using 1024 res and sample 20. And bog standard
antialiasing.
>>>>
>>>>So very BIG THANKS to all who replied!
>>>>
>>>>Joe Laffey wrote:
>>>>
>>>>
>>>>
>>>>>On Wed, 31 May 2006, Helen Bucknall wrote:
>>>>>
>>>>>
>>>>>
>>>>>>Is there a way to get shadow map shadows to stop
flickering even
>>>>>>when nothing is moving?
>>>>>>
>>>>>>This is what I have tried to no avail.
>>>>>>
>>>>>>On the spot light (cone angle less than 90):
>>>>>>Resolution up ( and down).
>>>>>>Softness up and down.
>>>>>>Samples up.
>>>>>>
>>>>>>None of these make any diference to the flickering.
>>>>>>
>>>>>>On the render panel:
>>>>>>I've played around with the antialiasing and jitter.
>>>>>>On the shadows page I've clicked off rebuild.
>>>>>>
>>>>>>
>>>>>
>>>>>If you view your scene from the light is there an
extremely large
>>>>>depth to the objects? In other words, are the closest
object to
>>>>>the light and the farthest object from the light extremely
far
>>>>>apart? This can lead to ZDepth (used for shadowmaps)
presision
>>>>>issues in some renderers.
>>>>>
>>>>>This is often the case if you have a huge ground plane.
>>>>>
>>>>>You may be able to exclude some objects, to help, or make
smaller
>>>>>versions that catch shadows.
>>>>>
>>>>>--
>>>>>Joe Laffey | Visual Effects for Film and
Video
>>>>>LAFFEY Computer Imaging |
-------------------------------------
>>>>>St. Louis, MO | Show Reel
http://LAFFEY.tv/?e01274
>>>>>USA
>>>>>
---
Unsubscribe? Mail Majordomo(at)Softimage.COM with the following text in body:
unsubscribe xsi
|
