E3 2011: The Witcher 2: Assassins of Kings Xbox 360 screens

I should have jumped in to this thread earlier, as your attempt to explain ubersampling has in fact described supersampled antialiasing instead. And also I'd like to point out that the X360's hardware actually supports supersampled antialiasing out of the box. The only problem is that it only works properly in a forward rendering pipeline, and since pretty much everyone went to a deferred pipeline just after the 360's release it doesn't get used all that much.

Assuming I'm reading the right techniques, I will now explain ubersampling.

Alright, so start with what we know about supersampled AA. It is the technique most commonly employed in the industry. You render more pixels than your output target has (2xSSAA renders twice as many pixels, 4xSSAA renders 4 times as many pixels, etc). This is then averaged out uniformly to the output target, usually by rendering a full screen quad with the SSAA target as the texture (thus letting hardware bilinear sampling do its job). The scene is rendered a total of once. The X360 games I worked on used this technique.

Ubersampling is a different beast. The trick to differentiating it is to not take the usual translation of the word uber - super - but to take another valid translation - over. Instead of sampling at a higher resolution, it samples at the target resolution multiple times. This is where it gets tricky, and why performance takes a hit. Rather than rendering from the same viewpoint each time, it very slightly alters the viewpoint. Rather than rendering, say, a 4x4 grid of pixels from one viewpoint, it renders that 4x4 grid from multiple viewpoints. It'll look similar, but because the viewpoint is slightly different it won't be the same. (Side note: this multiple viewpoint thing is how we used to take high res "screenshots" with the Wii.)

From there, it's a matter of rendering to the final target. You can't cheat and let the hardware do the work here. But cheating would defeat the point. Downsampling a normal SSAA target is basically a brute force average with no bias towards one pixel or the other. With ubersampling, it's entirely up to the shader writer as to how pixels get blended together. It could be as simple as a weighted average of all viewpoints, or you could even sample the texels around the current pixel and let that factor in to the equation. Or, more likely, signal processing is involved.

tl;dr - What pixel shaders did for rendering pixels, ubersampling does for antialiasing.