Analysis of techniques for improving the skylight

What are the best combinations of features for an optimal skylight?

To illustrate the points of the analysis, I have produced images that are designed to visually show the effects of the different techniques on the final image quality. Since this discussion is addressed to artists, I decided against more numerical or statistical results. The popular dictum that "an image is worth a thousand words" in even more obvious in this case anyway. 

Those images use standard available photoshop functions to implement image processing techniques. Each row represent the results from the setup described in the first column.

The first image is simply the portion of the image produced by the setup as it is output from the renderer. 

The second image is the same portion onto which the Edge Detection filter have been applied in order to enhance the noise present in the final image. 

The third image is the same portion but taken as the difference between the image and a reference image, averaged over 16 pixels and then multiplied by 12.8 in order to show the difference in shadow distribution relative to the reference image. The reference image was produced with a 148 lights skylight setup with width of 400cm and 7 rays each, oversampled and antialiased.

Increasing the number of lights

All lights are sun lights of 1cm width.

With low number (20) of sun lights of small width, marked shadow patterns appears on the ground, on the walls and on surrounding objects. Shadows are very unevenly distributed.
Increasing the number of lights (69) tends to decrease the banding of shadows and redistributes more evenly the shadows.
At higher number of lights (148), banding is less apparent but as revealed by the edge detect image, still visible. Also there is still uneven distribution of shadows as revealed by the spotty difference image.

Increasing the width of sun lights

148 sun lights of 500cm width. Consider that the skylight dome itself have a radius of 500cm.

Banding is much less apparent although still there as revealed by the edge detect image. Shadows are still unevenly distributed but the biggest disadvantage is the increased luminosity of the scene especially where there should be shadows.

Increasing the width of soft shadows lights

Soft shadow lights are large bulb or klieg lights that cast more than 1 ray. To help show the effect of increasing the width of the lights used in a skylight. I did several renders with the 20 bulb lights skylight with each light casting 2 rays. I then successively doubled the width of the light from 100cm to 1600cm. The skylight rig is 500cm in radius. It became obvious that simply increasing the width is not the perfect solution as it also increase the noise in the rendered image. In fact there is a compromise to achieve since increasing the width also increase the noise but increase the evenness of the shadows distribution.

100cm width. This is the less noisy image but the distribution is still uneven although much more blurry that without soft shadows.
200cm width. Noise is quite visible and shadows are still unevenly distributed.
400cm width. This seems to be the best compromise setup in term of even distribution of shadows vs quantity of noise. narrower, we get uneven distribution of shadows and wider we get more noise.
800cm width. Even with such a large light, the distribution of shadows do not get much more even. However the noise is significantly increased.

The same need for a compromise can be observed with a 69 light skylight setup with 2 rays per lights although increasing the number of lights generally produces less noisy images with more evenly distributed shadows.

100cm width. Noise is almost unnoticeable although still there as revealed by the edge detect image. The difference image clearly shows spots on uneven distribution of lights.
200cm width. Noise is still at an acceptable level but the shadow distribution is still uneven.
400cm width. Noise start to become annoying but the shadows distribution is quite even.

Increasing the number of rays per lights

For the next tests, the 69 lights skylight was used with a width of 300cm for the lights. From the test shown above. 300cm seemed to be the best compromise.

3 rays per light for a total of 207 rays. Not a big improvement over the 2 ray setup. The noise is barely noticeable although present.
7 rays per lights for a total of 483 rays.

Subjective perception of noise vs banding

Basically the difference in render quality between a 483 lights skylight and a 69 lights skylight with 7 rays each (483 rays) shouldn't be expected to be very different. One will produce almost unnoticeable banding while the other will produce almost unnoticeable noise. However. there are scenes where the banding will show up even with very high number of lights. For instance, it is easy to place a square patch parallel to the ground and at the appropriate height from the ground that will show banding even with 1000 lights.

The main disadvantage of banding from a subjective point of view is that the pattern they produce are structured. This structure makes the patterns very easy to spot visually. Noise on the other hand is unstructured and produces no perceivable pattern.

Banding becomes even more noticeable during animation as the pattern moves on objects and makes a very busy animation. However, given the large render times that results from the large amount of light rays to be computed, animation might not be the most suitable application for skylights.

Our visual system seems to be at ease with noise. We are not disturbed by film grain in cinema or by noise in scanlines on TV or even by noise in badly compressed movie files. The eye have the ability to integrate the noise over time and just forget about it.

Render Times

Here are a few render times (with average render time per ray in the 3rd column). The images where rendered on a Pentium III, 500mHz with 256 megs of RAM. The image is 600 x 320 and no oversampling or antialiased where used unless specifically noted otherwise. It is a modified Toys project from the Hash CD-ROM, frame 0:03:28.