drohne said:
b. the second description of 'temporal AA' pasted in here describes the effect as it's implemented on newish ati pc graphics cards -- the ps3 version of dmc4 is just blending frames with jittered versions of preceding frames
c. my god it's difficult to establish a simple fact on gaf. i think i give up
I think that there may be some misunderstanding regarding my previous technical posts, which i thought were quite logical and simple even for the layman, but perhaps there were some ambiguities.
Some of you clearly are very interested in the effect and its more logical description rather than just relying on names and jargon which mean little, so i supplied some of my time to explain the concepts and some key ideas and what I know, which can be interesting for people interested in image quality problems and for the technologically inclined.
To the rest who are not so interested, pls skip this post.
okay,to get this straight, I will explain aliasing properly from the ground up.
Television Pixels and Jagged Edges
Television pixels are made up of squares. real life Images are not, and have a a smooth continuous representation.
Now, computers' brain (ie, the microchip) cannot imagine something continuous. To compensate, they calculate the appropriate colours only at certain points, usually at the corner of one of the squares of pixels, and the entire pixel would be lit up according to the colour at a single corner.
This has problems, because sometimes, certain objects, especially far away ones, are
smaller than the size of a pixel, so you have (in the real image) colour variation across the size of the pixel, which is not captured. Hence, certain tiny details appear as jarring blocks. These are what you call
aliasing.
Removing the Jagged Edge Conventionally
The usual methodology to solve this is to take more samples. What this means is that the computer will choose two different points across the pixel and calculate their colours and average them out. This outputs a colour that is the average of 2 points rather than just at a single corner, which is slightly more accurate. There wouldn't be additional detail, as the pixel is still lighted up using only one colour. But the transition from adjacent pixels would appear smoother. This, logically, is usually done at 2 opposite corners of the square pixel. This is what you would call 2x anti aliasing. 4x simply denotes that the computer would calculate at 4 corners of the square pixel instead.
As you would therefore notice, 2x anti-aliasing would mean that the computer would have to calculate the colours at 2 times more points, and 4x AA would mean 4x more points. This means a lot more information to process and store, which is why the process can be expensive. Usual AA techniques are simply slight variations of this greater principle.
Another Method - Temporal Antialiasing
now, temporal anti aliasing employs a slightly different technique. It uses the fact that our eyes tend to merge images that are 1/12 of a second apart. Contrary to popular belief, this does not imply we cannot observe increased fluidity at higher frames, it simply implies that the eyes naturally blend detail from the preceeding frame to the next, so some details are carried over. This is called visual inertia, and is the reason why we don't see out televisions as scan lines but full images.
now the whole idea of AA is to sample the colour at 2 corners of the pixel. Since at 60 frames per second, the images are only 1/60s apart which is 5 times less than 1/12 of a second, we can just calculate the colour at one corner for 1 frame, and in the very next frame, we calculate the colour at the other corner. Our eyes would mix the 2 images and and produce an average automatically as 2 different colours are flashed almost simultaneously. But this is exactly the same effect as 2xAA!! And this is exactly the same effect introduced in ATI card to improve image quality.
I am sure you have experienced this before. If you have a wheel of red, green and blue, spin it fast enough and it would look white!!!
To illustrate, look at this picture:
now, this is from ATI, so they assume you would already have 2xAA minimum enabled. At 2xAA, one frame would have 2 corner's colour calculated and averaged, and in the next frame, the other 2 corner's colours would also be calculated. Combined, the effect would be similar to 4xAA. On the PS3, what they do is to skip the AA part and just calculate one corner in one frame, and the other corner in the next frame. So you get 2xAA instead of the 4xAA in the diagram above. It is that simple.
Motion and Image
Now, you all now that the effect used above works great for static scenes. What about motion? well, the idea is that in motion scenes, the colour calculated for in the previous frame would have moved to another pixel already before the next frame. This means that the colour would blend in with the colour in the new frame even though it isn't there anymore! But wait! If you consider that even if you emply standard 2xAA instead of TAA, in motion the colours in those images would ALSO have moved while the next frame is beling calculated! That means your eyes are blending together and produces 2 averages of colour instead, since the 2xAA image is already averaged out for you. But the averages usually do not differ by much even if you average again. It is hard to explain this, but the net result should be an almost imperceptible difference in the viewed image when on a television screen. In any case, the eyes cannot perceive detail well for objects in motion (you just try to catch the number plate of a ferrari).
Conclusion? In motion, they would look almost the same. Which is why ATI includes it in its software - because it improves overall image quality.
What they did in DMC4
Well, the technique they used is just as i described above, and it is also the same idea for the option in ATI's videocards. One point of contention would be that for the PS3, the folks at B3D claim that the frames are actually blended (or averaged out) by the processors themselves, which is strange, and the concept above does not clearly indicate a need to do so. ATI themselves do not appear to have employed this extra step, from what I know.The net result of this is that the colours would carry over a few frames into the future. For example, in the 1st frame, lets say, yellow is calculated, in the 2nd frame, brown is calculated, and the average is yellow-brown, the 3rd frame, say green, would mix with yellow-brown and average out to green-yellow-brown, and this would carry on for a few frames.
The net result of this could be a streaking effect or motion blur. This may, or may not be intentional, as streaks tend to represent speed in images. This, however, is easily avoidable simply by switching off the blending of frames.