Starting from a given frame, he illustrates the different stages in the rendering: creation of the G buffer, shadow map, ambient occlusion, light prepass, how opaque and transparent objects are treated differently, volumetric lights, bloom effect in LDR, anti-aliasing and color correction, the depth of field, and finally the object interaction visual feedback.
Here are a few screenshots stolen from his article:
Rendering from compressed textures, Beers et al., proceedings of SIGGRAPH 1996 “This one (out of 3) of the 1st texture compression papers ever! Uses VQ so probably not something you want today, but major eye opener!”
Shadow mapping is a popular way of getting dynamic shadows, but suffers from aliasing artifacts that cannot be addressed by usual texture filtering. The reason boils down to the fact that the average of depth test results (which is what we want) is not the same as the result of a test on the average of depths (which is what hardware does).
The trivial way to do it anyway is the Percentage Closer Filtering (PCF) technique, and usually stands in papers as the expensive upper bound.
Variance Shadow Maps are a simple technique allowing filtering, including some Gaussian blur for example, thus giving soft shadows (the blur does not depends on the distance from the occluder though). The main drawback of the algorithm is the light bleeding artifact that occurs as soon as the scene complexity is too important. I also found it to be fairly expensive in terms of texture memory since it requires twice as much as regular shadow maps, and another times two for blurring.
The Chapter 8 of the GPU Gems 3 book gives crystal clear explanations of the technique and its implementation, as well as its drawbacks and how to address them (which quite often missing from publications)