In|Framez® Papers

Depth Tinting Analysis

By Homam Bahnassi

Even though XSI6's out-of-the-box hair system can be quite fast and stable in rendering, in some cases it can give considerable rendering times for frames with hundreds of thousands of hair strands with complex shading/shadowing options. For that, I spent some time figuring out different approaches for rendering hair in XSI including my own approach which I call “Depth Tinting”. Below are the results of the comparison between the possible techniques, in addition to rendered captures that show the rendering quality differences between the methods.

The Test

The test machine was chosen with quite-low configuration to emphasize the rendering time and show clear differences between the techniques. The test machine's specs are:

CPU AMD 2000+
RAM 1GB
OS WinXP SP2
XSI Version 6.0

And these are the rendering settings used for the test:

AA Min	-1
AA Max	1
AA Sample filter type	Triangle
AA Sample filter size	2
BSP Max Depth	45
BSP Max Leaf Size	10
ShadowMap Resolution	1024
ShadowMap Sampling	24
Volumic ShadowMap Sub-Sampling	2
AreaLight U Sampling	3
AreaLight V Sampling	3

The Graphs

The following graphs show the rendering time for each method with different hair count. You’ll notice how “Depth Tinting” is an all-case winner. The rendering results produced are quite acceptable in comparison to the other methods with longer rendering times.

Hair Count (strands) Ray-traced Area Light Shadow Map Depth Tinting

4000 127.11 704.56 67.84 45.29

8000 274.66 1395.26 99.13 62.24

12000 345.1 2149.26 118.35 74.36

16000 421.93 2400.71 138.23 78.98

20000 491.35 2926.08 159.69 85

In addition, we compute the percentage between each rendering technique and Depth Tinting. Notice that rendering more hair strands makes Depth Tinting more effective. The difference is clarified by computing the percentage between the rendering time for 20000 strands to 4000 strands:

Hair Count (strands) Ray-traced % Area Light % Shadow Map %

4000 280.65798 1555.66 149.790241

8000 441.29177 2241.74 159.270566

12000 464.0936 2890.34 159.15815

16000 534.22385 3039.64 175.018992

20000 578.05882 3442.45 187.870588

(20000/4000) % 2.0596557 2.21285 1.25422449

Hair Count (strands)	Ray-traced	Area Light	Shadow Map	Depth Tinting
4000	127.11	704.56	67.84	45.29
8000	274.66	1395.26	99.13	62.24
12000	345.1	2149.26	118.35	74.36
16000	421.93	2400.71	138.23	78.98
20000	491.35	2926.08	159.69	85

Hair Count (strands)	Ray-traced %	Area Light %	Shadow Map %
4000	280.65798	1555.66	149.790241
8000	441.29177	2241.74	159.270566
12000	464.0936	2890.34	159.15815
16000	534.22385	3039.64	175.018992
20000	578.05882	3442.45	187.870588
(20000/4000) %	2.0596557	2.21285	1.25422449

Limitations

Although this technique has potential as a rendering time/quality balance, it still has limitations at the moment, which prevent it from being used widely in rendering hair in XSI or any similar package that depends on the “Hair & Shave Cut” hair system.
These limitations can be listed as the following…

Rendering hard crisp shadows can give inaccurate/approximated shadows; thus it’s limited to rendering soft/blurred shadows.
Hard to get the same quality for very long hair (Pony tails,…) as opposed to other standard techniques (Shadow Maps, Area Lights,…).
The current algorithm can fail generally in predicting the correct values for mimicking the rich results from other techniques (e.g. Area Lights), this is because it depends a lot on artistic sense in modifying the final values. Thus the technique is not fully automated at the moment and requires artist modifications.

This technique is still a prototype, the main algorithm at the moment is based on experimental mathematical equations for predicting hair & surface attributes. Replacing these equations with more clever equivalents is our next high-priority task.

Demo

For viewing an actual test drive of this technique, you can download the Querks demo that is based completely on this technique, each frame took about 45 secs on the mentioned test hardware and rendering settings.