I found the right hand image somewhere in the internet on a site about global illumination rendering.
In my opinion it was produced by some kind of density estimation algorithm which works on
polygon level. Because this scene shows nearly all the interesting effects like caustics,
direct/indirect/diffuse illumination, area light, soft shadows, specular reflection and refraction,
color bleeding, etc. it is a very good test-scene for comparison of different rendering algorithms.
I adapted the scene for WinOSi and rendered it with the iterative-two-pass-optical-simulation engine
which works on pixel level. Surprisingly there are some characteristic differences between these two
global illumination renderers:
1) In the WinOSi image, there is a small bright light point in the middle
of the red wall, and on the right side of the glass sphere, which is completely missing on the right
2) The color bleeding seems stronger in the right hand image, while
the WinOSi image has relative weak colors considering the walls material consisted of (asked for)
pure red and blue.
3) The shadow under the mirror sphere is darker and has sharper edges
in the WinOSI image.
4) The shape of the caustic maintains the same rectangular shape as the lamp
in the WinOSi image, while it's circular in the right hand image.
5) In the right hand image, the mirror-sphere shows some bright light points
on the floor and on the walls, which are not visible on the walls directly, nor on the WinOSi
6) Rendering time for the WinOSI image was over an hour (1 GHz computer) and
the right hand image was probably rendered within a few minutes.
I think that some of the inaccuracies are the price you have to pay for estimating the illumination
-function on a polygon mesh of too low resolution, instead of true per pixel evaluation. Some other
differences may actually be a design- or implementation-defect in WinOSi or the other renderer.
Perhaps we have to recreate this scene in the real world and then take a photo of it to determine
which image is correct.