spot blog: Filter Differences between FLAM3 and Apophysis

March 16, 2005

Filter Differences between FLAM3 and Apophysis

A few days ago I got a bug report on FLAM3 from Robert Joy. He had done a render with FLAM3 and with Apophysis, and the FLAM3 render came out noisier.

Today I had a chance to look into it. I rendered 8 images using a genome Heaven 20 from Robert. I ran 4 different filter settings with each of FLAM3 and Apophysis.

The first pair with oversample 3 and filter radius of 0.2 shows the difference reported by Robert: the FLAM3 image is grainier.
In the second pair with oversample 3 and filter 1.0 both have low noise level, but the FLAM3 image is a bit sharper.
In the third pair with oversample 1 and filter 1.0 have similar noise levels, but again the FLAM3 image is sharper. Comparing them with the 2nd pair you can see the effect of the oversampling in reducing the staircase on the bright white curves.
The last pair has oversample 1 and filter 0.2 and they look nearly identical.

So what's happening? The fast answer is yes they differ but Apophysis is wrong. The filter setting of 0.2 is so narrow that with oversample 3 it amounts to using only the center one of the nine subpixels. Ignoring most of your samples makes for more noise. The solution: use a larger filter!

Why doesn't Apophysis have the same result? I see 3 differences in the filter code:

When the FLAM3 code was transliterated from C into Pascal to create Apophysis, a change was introduced into how the filter is calculated. Instead of being truncated to an integer it was rounded. Compare the definition of filter_width in Apophysis and an old version of FLAM3. This results in Apophysis sometimes using larger filters and hence producing blurrier images.
FILTER_CUTOFF changed from 2.5 to 1.8 in v1.10 on 3/4/2004 and this change wasn't picked up by Apophysis. This should be just a performance optimization (smaller filters with less zeroes).
v1.13 on 3/26/2004 fixed the deficiency where the size of the filter was truncated to an integer number of subpixels before the filter weights are calculated. Apophysis didn't pick this one up either.

The bottom line: of all 8 images the one that looks best was rendered with FLAM3. It also completed about 2x faster than Apophysis. Our long-range plan is to ditch the apo renderer and have it call a FLAM3 library. In the meantime it wouldn't be hard to update Apophysis and improve its results. Though I must admit the more I think about it the more I believe the Right Thing is to rewrite Apophysis in python...

I think it would be reasonable for FLAM3 to print a warning when the filter radius is "too small". And perhaps to enforce a minimum so that information is not lost. I'll think about this. Posted by spot at March 16, 2005 12:57 AM

Comments

Thanks Scott for checking into this. I really appreciate your effort. Your explanation makes sense and provides insight into why there are differences. Is it safe to assume that sustantial increases to FR would be a reasonable workaround for this issue (at least with larger oversample values)? You elude to that but don't make a definitive statement about it. If so, would the FR need to be proportionally increased as the image is scaled to larger sizes? Thanks again, Robert Joy Posted by: Robert Joy at March 16, 2005 09:41 AM

well, that sounds like fun scott, i just hope that if you guys do decide to ditch the internal renderer, you make flam3 at least accept the limitvibrancy false script. just a hope. Posted by: james at March 16, 2005 09:53 AM

Robert: no the filter radius should not be scaled with the image size. Just don't use such a small value. And don't expect the filtering from apophysis and flam3 to match. Posted by: spot at March 16, 2005 12:16 PM