Remember the artifact visible on the main page - that happens when the image rendered for each patch is too small, and a group of light patches get exactly the same number of very bright pixels from the outside. Then the group of patches next to them which get one more bright pixel appears significantly brighter. You need to find correct size of light patch vs. image rendered for each patch.
Below is an image rendered for one patch. This one is huge - 200 by 200 pixels. It would take ages to render even the simplest scene. You can get reasonably good results even with a 10 by 10 image.
A few points to note:
Having done a few passes of radiosity, I switched off texturing and rendered the image below. OK, we knew anyway that things look correct, just because of the brightness slowly fading away the further we are from the opening. But on this image you can see that the green wall is reflecting quite a lot of green light. It's not as pronounced in the textured image, but if you look carefully you'll see it.
Just see how much realism good lighting adds to the scene. Below is an image of the same scene with lightmap switched off. OK, there are other ways to have a very realistic scene, and the scene I use in this example isn't particularly realistic anyway, but still the difference is enormous.
In the image below you can very clearly see the separate light patches. Human eye is extremely good at seeing edges, so bilinear filtering really helps.
Never mind a small bug in my bilinear filtering which causes two light patches to show in a very dark color. Can't be bothered to fix that.
I realise that there is a problem with clipping floating point colors, and I even realise that it's not too hard to implement an exposure function, but I don't really feel like doing that, not in the least because that would mean rendering all the images for the web again.