Henrick has written a new variation on the FSAA testing routine, here's what he says about it:

"The red/green part is two intersecting quads. This makes the edge harder to detect for dynamic FSAA.

Now there is also a Geometry mode where the texture is replaced with a tiny white quad for each white texel. This means one quad for each pixel. In geometry mode, pressing "Q" and "A" will increase/decrease a "dithering factor" that makes the program skip different amount of quads. This can reduce the stress on framebuffer size, but will make the dots look dithered. Each time the screen is re-rendered, a new set of quads are rendered. So forcing the program to re-render the screen will loop through all positions for the quads."

Lets take a look at the output from this:

If we zoom in on the red/green area we can see that Wildcat III has no issues with the rendering of the quads:

Once again we can see that the Wildcat has rendered the red/green area with a full set of 15 gradients. (Note: again these are gif images, download the originals if you wish look at them in more detail)

What's more interesting is what happens when we use the geometry mode on the application, and increase the 'dithering factor'. Here are a few close-ups of the output in that mode:

(Note: Again these are gif images and the white areas have been enhanced to display better. All images were taken from the same location on the output and zoomed 400%).

As we can see as the sampling pattern is changing from one pixel to the next and as the 'dithering factor' is increased we get an increasingly random sampling pattern. This is a good indication that Wildcat III is capable of truly stochastic FSAA sampling. However, considering the sampling patterns on the edges or the red/green area does not change there would still seem to be a level of 'intelligent' sampling as well.

The original images of these outputs can be downloaded here and the new FSAA test application here.