Anim8or Community

General Category => Anim8or v0.98 Discussion Forum => Topic started by: CobraSpectre on March 20, 2008, 09:23:47 pm

Title: ART Light Attributes and AA
Post by: CobraSpectre on March 20, 2008, 09:23:47 pm
I was testing out the new light attributes in ART. The jittered and multijittered lights look like monte carlo even if the lights are set to ordered. Regular looks like ordered. Edit: Sometimes (with the right settings) regular can look similar to ordered but they don't look the same.

The other problem is that a light's shadow softness and noise can look very different between ART and scanline.

Can we have more details as to how the new AA options compare? And how the samples work and how they are different than scanline?

The new Movie Image is nice and works well.
Title: Re: ART Light Attributes and AA
Post by: CobraSpectre on March 20, 2008, 11:15:29 pm
After some more tests...

Regular doesn't really look like ordered. It can be similar, but requires careful settings. On the buckyball example it looks like a low sample ordered, but with noise.

Light shadow softness and AA samples are closely tied together. To get smoother noise you need much higher samples.

Min Max samples appear to be ignored as do monte carlo vs ordered.

Z-Near settings are ignored.
Title: Re: ART Light Attributes and AA
Post by: Steve on March 21, 2008, 05:15:10 am
Antialiasing is done iquite differently in the ray tracer and the scanline renderers.

The scanline renderer always uses a 4x4 grid of samples per pixel with all the samples in 4 rows and 4 columns.  This is why it's so fast - many of the computations are done once per image instead of per pixel or per sample.  But even samples are very bad for soft shadows.  So for soft shadows multiple shadow samples are takes for each image sample.  First a few samples are taken.  If any of them are different (i.e. one is in a shadow and one is not) then that pixel and the neighboring ones are samples a higher number of times.  This allows a high number of samples to be used on shadow edges and fewer in areas that they aren't needed because they're all the same.

The ray tracer uses just one shadow sample per image sample.  But because each sample is independently traced a better distribution can be done.  Multijitrered is a very good method - much better than random or Monte Carlo.  Likewise glossy surfaces trace just one sample per pixel.  That's why non-aa glossy surfaces are so speckled.  Over all it's best to do it this way because you sample all the varying parameters evenly.  The scanline renderer samples shadows at a much higher rate that it does the image - but that's required due to very thing that makes scanline do fast - the regularly spaced image samples.

So you will not be able to exactly match scanline images with ray traced for this an a dozen other reasons.

You're right.  There is no near and far Z with ray tracing.  There is no need for it becuase thereisn't a perspective projection, so no divide and all of it's related precision problems.  Everything is done in world coordinates.