Comment by GuB-42
14 days ago
> just mess with it and see what happens
And even if you know every detail, that's still the best course of action, I think. Which kind of antialiasing you prefer, and how it trades with performance and resolution is highly subjective, and it can be "none".
There are 3 components to rescaling/rendering pixels: aliasing, sharpness and locality. Aliasing is, well, aliasing, sharpness is the opposite of blurriness, and locality is about these "ringing" artefacts you often see in highly compressed images and videos. You can't be perfect on all three. Disabling antialiasing gives you the sharpest image with no ringing artefacts, but you get these ugly staircase effects. Typical antialiasing trades this for blurriness, in fact, FXAA is literally a (selective) blur, that's why some people don't like it. More advanced algorithms can give you both antialiasing and sharpness, but you will get these ringing artefacts. The best of course is to increase resolution until none of these effects become noticeable, but you need the hardware.
The best algorithms attempt to find a good looking balance between all these factors and performance, but "good looking" is subjective, that's why your best bet is to try for yourself. Or just keep the defaults, as it is likely to be set to what the majority of the people prefer.
Oh there are lots more axes than just the 3 aliasing, sharpness, and locality ones. Those are the main tradeoffs for a pixel sampling or convolution filter choice, when downscaling an image, say between nearest neighbor vs bilinear vs Mitchell (bicubic). But the antialiasing methods in this article have many tradeoffs that aren’t on the aliasing-sharpness-locality spectrum. Other issues with real time AA methods include bias, correctness, noise, quality, temporal effects, compositing/blending issues, etc. And the topic gets much wider when we start talking about DLSS, we don’t even have established terminology for the many different kinds of tradeoffs neural networks give us. Anyway just noting that the main highlights of discussion in the article, which are MSAA and AAA (and references to TAA and others), don’t fit in the aliasing-sharpness-locality space. MSAA’s tradeoffs include it only running on geometry & texture edges, and its ‘wrong order’ samples or ‘double edges’ noted in the article. TAA has a temporal aspect and is most known for ghosting. AAA as described here doesn’t necessarily blend correctly and in general it can’t handle multiple arbitrary sub-pixel events, it really only works well if there’s one edge crossing a pixel.