Ive been noticing that with a default mantra render that there are some aliasing especially with diagonal straight edges. Cranked up the super samples to 8 but it still didn't relieve the problem.

I looked at maya renders and I noticed that it uses triangular based filtering. I tried the barlette 2x2 filter in houdini and it really helped a lot! Mental ray uses gausian 3x3…but that didnt look as nice as the barlette 2x2


Wondering if anyone has tips for settings that achieve the best balance between detail and smoothness in mantra?

Hmm, I oughta try that. Switching between the default filter and catmull-rom I didn't really notice much of a difference. But speaking of AA, it might be nice if mantra implemented adaptive anti-aliasing ala Brazil. So it'll put down more samples in “noisier” areas of the image and fewer in the clean parts.

Wren, there is 3 parameters for AA: super sampling, jitter and AA filter. Try to set jitter to zero (if it's a still image) and play with different filter types. Maybe set supersample 4x4. It is possible to find the compromise

DaJuice, is it same to

This option (-v option on the command line to mantra) can be used to specify a different anti-aliasing mechanism. Variance Anti-aliasing takes an argument which specifies the acceptable tolerance in color contrast between pixels. If the tolerance is not met, anti-aliasing is performed. This will anti-alias areas of the image which tend to have aliasing (i.e. edges of primitives, texture maps). Increasing the value of the variance (maximum of 1) results in faster renderings. Decreasing the value (min- imum of 0) produces higher quality images.

The -v options sounds like lightwave's adaptive AA where it renders the image once and then it looks for areas of contrast and samples it more in areas that are higher than the contrast setting.

This saves render time however it means that you are only getting one sample in the below threshold areas. For quality renders, a minimum of 3 samples for the whole image is needed. (especially using motion blur)

It would be useful if the minimum number of samples could be controlled along with the maximum.

ps. when I was referring to barlett 2x2 im talking about the width of the barlette filter. I found that 5x5 to 7x7 produce clean results for super samples.

We noticed the same thing here - it seems as if a Gaussian 3.5x3.5 in Mantra looks similar to a PrMan Gaussian 2x2 - at least with straight geometric aliasing issues. I raised this with Jeff Wagner just last week when he visited LA.

Im really feeling conflicted here because the barlette 2x2 filter does such a good job cleaning up the edges. However I noticed the whole image loses its sharpness, not just those jaggie edges. I suppose it would be possible to render the image at double the resolution with the barlet filter then in comp scale it down so it contains the detail and the softness together. …. but im an apprentice user …. so thats not an option for me.

There are some technical issues which are interesting, as Jason mentions, but unless you're doing some extremely serious, sharp, detailed rendering, it's not a big deal to find a decent balance between detail and edges without doing that oversized render thing. Besides - it needs to be looked at on a project by project basis - once you get the image in there, moving along with the camera and with mblur - is it a problem? Just staring at a still image of a render or a 32 bit probe image will only get you so far outside of theory, which is definitely relevant, but not the end of it all.

We've always found the default settings with mantra to lean a bit too much towards speed rather than image quality. Upping the SS goes a long way to addressing the edge detail. The rest of it - pixel filtering - will always be a mix of curse and blessing. The adaptive thing sounds cool. The look of mblur - don't get me started on patents…

Cheers,

J.C.

The thing with the default setting for mantra is that it gives harsh geometric edges. It seems to me that the Gaussian filter it implements is not as heavy as prmans, but I do prefer prmans default look. There is no real speed hit for a larger kernel at these sizes and I'd prefer it if either the default is changed, or the Gaussian kernel is. I believe the filters are global across Houdini so perhaps the default needs changing.