» Tue May 29, 2012 4:08 am
This method is generally referred to as downsampled supersampling AA because you natively render the image at a higher internal resolution and then scale it down to a lower target resolution. The end result is that the higher number of pixel data samples used to create a single pixel are averaged together to create a more accurate pixel color that has less contrast/aliasing with adjacent pixels. For example instead of black or white, you will get a shade of grey, allowing a less aliased transition between the two colors.
The key benefit is that supersampling AA will apply AA to the entire scene (sometimes referred to as full-scene AA), unlike MSAA which only covers geometry and TSAA that only covers transparencies/alpha textures. Supersampling will apply AA to everything in the scene including textures and shaders. Beyond the lack of jaggies, supersampling AA in motion has a calming effect on the scene as it reduces the 'pixel crawling' phenomena you may see when high constrast pixels switch back and forth while moving.
The key downsides are that any supersampling is extremely expensive performance-wise, since you are generally rendering the entire image at a higher resolution or # of samples. The performance impact is not quite linear by number of samples but it is heavy; generally you will see a much bigger requirement in terms of GPU horsepower, VRAM usage, and required bandwidth. Another key downside is that because AA applies to the entire scene including textures and ultimately uses an averaging/blending algorithm to combine multiple pixels into one, you can get a blurring effect on the entire image compared to traditional MSAA+TSAA methods. There are some tricks to reduce the impact of blurring, like reducing texture LOD bias, but this can be problematic for some games.
I have not looked into the downsampling methods out there for some time, only for GTA4 using Boris' ENBseries mod for that game because no other form of AA was available. But that's a good game to look into if you want to see the readily apparent results. The IQ is much improved but the performance impact is extremely steep and the game was virtually unplayable on some cards that didn't have enough VRAM as a result of the tremendous increase in VRAM required. FFXI is another game where players used a registry mod to have the game render at a higher resolution and then scale itself down to native target resolution.
Personally, I don't think this method is worth it for a game like Skyrim that already looks good (with mods) and provides a wide array of AA modes already. The biggest concern to me is the amount of VRAM Skyrim uses with just texture mods, that's going to skyrocket once you enable supersampling. If you do want to experiment with higher than available AA modes, you should look into MSAA+TSAA+FXAA first, then if you want to notch it up from there, you can use the unofficial SGSSAA modes available via Nvidia Inspector.
SGSSAA (Sparse Grid) is another form of supersampling AA that builds on the fundamentals of MSAA and TSAA, but instead of rendering at a higher resolution, it uses sub-pixel samples to sample each pixel numerous times using a sparse grid array and then composites those pixels into a single pixel. Basically, it applies the benefits of transparency supersampled AA to the entire scene instead of just to the alpha textures in the scene. It also carries a heavy performane hit but I do believe its a bit lighter overall then downsampling because it doesn't require each intermediate buffer and the output framebuffers to be rendered at a higher internal resolution.
I'm personally a big fan of FXAA because while it does blur the scene, the performance hit and increase in VRAM are negligible and it does an amazing job of applying FSAA. So in summary, I'd play with em all to see the differences and what you like, then settle on something with MSAA + TSAA + FXAA to get a similar FSAA effect without the huge performance hit.