Indigo image quality.

[tungee]

Hi ono,which image sampling you use for Indigo?
is it gaussian?, i heard that mitchell-netravali is much more better than gaussian, this means much sharper without aliasing.
in some cases the indigo images looks bit blured with and without the post.

[OnoSendai]

Indigo uses a Gaussian sampling filter.
Mitchell Netravali is a bit sharper, but it's a subjective matter which is better.
If you use a super sample factor > 1, the images will be pretty sharp even with a Gaussian filter.

[tungee]

Hi ONO,
would it be hard to implement the mitchell netravali filter in Indigo? i have seen some sunflow pics , they are really very sharp and crisp.

http://www.cs.cornell.edu/Courses/cs465 ... tes-v2.pdf


http://www.blackwell-synergy.com/doi/pd ... 06.00919.x

[tinman999]

IMO mitchell netravali filter is not good for filtering global illumination image, and indeed any filter which uses negative weights unless you apply the filter after you tonemapped your global illumination image. This is because the negative weights of the mitchell netravali can cause black edges (for example at the edges of light source) this dues to float point rounding error.

[CTZn]

IMO mitchell netravali filter is not good for filtering global illumination image, and indeed any filter which uses negative weights unless you apply the filter after you tonemapped your global illumination image. This is because the negative weights of the mitchell netravali can cause black edges (for example at the edges of light source) this dues to float point rounding error.

I agree a priori, this was the cause of a sampling issue in mental ray with Final Gather, known as the "black spots" issue IIRC.

What about elliptical filtering ? It's quite renowned I believe...

[IanT]

IMO mitchell netravali filter is not good for filtering global illumination image, and indeed any filter which uses negative weights unless you apply the filter after you tonemapped your global illumination image. This is because the negative weights of the mitchell netravali can cause black edges (for example at the edges of light source) this dues to float point rounding error.

Exactly matches my own experience. Supersampling can actually reduce the "black edge" effect of the MN filter but it still doesn't totally disappear

Ian.

[zsouthboy]

I'm going to vote for simple gaussian, too.

If your results seem less sharp.. sharpen 'em. No reason to cause every render to have a higher relative error, for some sharpness.

[Kram1032]

It really doesn't matter, to me, as long, as it wont entirely slows indigo down and/or there is a reasonable effect.

[tungee]

what about catmull-rom Filter?
I dont like to postsharp because of antialiasing.
thanx for da replies

[IanT]

what about catmull-rom Filter?

That has a negative lobe too sadly It's also bloody slow to evaluate (even slower than MN)

Ian.

[zsouthboy]

If you're seeing aliasing when you do post-sharpening, you're doing it wrong.

[IanT]

If you're seeing aliasing when you do post-sharpening, you're doing it wrong.

Yes, good point. Try decomposing into HSV or equivalent, sharpen only the luminance channel and then recompose.

Sharpening can make aliasing worse but only if you have it there in the first place. Using Indigo's supersampling at render time will prevent that.

Ian.

[Xman]

For building render,the Gaussian sampling it's OK.
but for industry design render,sharpen sampling is necessary.

PS.in fact, I much more hope indigo can own two different samples.

[CTZn]

PS.in fact, I much more hope indigo can own two different samples.

That would make sense. For instance, mental ray has five sampling filters:

Box, Triangle, Gaussian, Mitchell, Lanczos. The two latters are clipped to avoid negative values.

Because “plain” Lanczos and Mitchell may produce negative values the new filter types are “clipped” variants to ensure positive values. The filtered result samples are clipped to the min/max range of input samples. The final pixel in the image will therefore not contain any out-of-range values _produced by the filter_, as might be the case for regular Mitchell and Lanczos filters.

Additionally, the filter size is not fixed but set by hand (has default value for each type).

[IanT]

That's a strange way of doing it ... it's exactly the negative lobe of these filters that gives them their edge-enhancement properties

Ian.