Nearest neighborhood upscaling + 8x MSAA (Multisample antialiasing): Would be good?

[luquinhas0021]

Everyone hear talks that Point upscaling is lossless. But it makes a lot of aliasing in image. I've researched and discovered that 8x MSAA is a so far better option for antialiasing. For Point upscaling, would maintain the lossless approach? Point upscaling more 8x MSAA would make no aliasing images and, simultaneously, kept all details from LR image?

[captainadamo]

How many of these threads do you actually need?

[luquinhas0021]

What? Please, if you know about, answer my ask.

[foxyshadis]

Multisampling is for reducing an image, and therefore it's also lossy (although it's one of the better-looking fast ways to reduce an image). How exactly do you propose to multisample a digital image for upscaling? You'd have to use a resampling algorithm (nearest, bilinear, spline, lanczos, etc) to generate the sample, at which point you may as well use the resampling algorithm directly and save yourself a lot of time.

You still haven't explained why lossless is necessary. Visually pleasing or easily studied should be most important.

[luquinhas0021]

foxyshadis, msaa is a antialiasing technique, and makes downscaling after makes upscaling, and it is only done in edges, instead of most antialiasing filters. The idea is make upscaling with pointresize, then adds 8x msaa, for reduce or disappear with aliasing caused by nn resize. Or, better, make a algorithm that merges this two approaches. What do you think?

[foxyshadis]

foxyshadis, msaa is a antialiasing technique, and makes downscaling after makes upscaling, and it is only done in edges, instead of most antialiasing filters. The idea is make upscaling with pointresize, then adds 8x msaa, for reduce or disappear with aliasing caused by nn resize. Or, better, make a algorithm that merges this two approaches. What do you think?

Antialiasing is fancy downscaling! In a 3D scene, fractals, and other CGI, you can render infinitely large and edges will always have more detail, so the more samples you use the more detail you retain. MSAA does little to nothing for textures, which are 2D images, same as the images and video we're talking about. You can't render an already rendered picture any larger, and that's why research into ways to upscale is so difficult and slow, whereas innovations in AA all come from choosing sampling locations differently.

You should test the result of MSAA from a point resize. You can easily do it: Point resize up to something huge, then down from 8 offsets less than your scale factor, then merge all 8 results with average(). That's the MSAA algorithm. Guess what difference you'll see? Nothing. Sampling and averaging the same color 8 times still gives you the same color.

[Bloax]

MSAA actually doesn't do anything for textures, since the reason it is faster than supersampling (the oldest AA method) aka SSAA/FSAA/whatever in the first place is because it is explicitly targetting polygon edges.
Which is also why transparent textures on models remain aliased as fuck. (Supersampling looks really great on textures if you kill the bi/trilinear filtering, by the way.)

This also means that MSAA is entirely useless for video - you can't render polygon edges at a higher resolution and downscale them in a video, simply because the only polygon edges in a video are the edges of the surface the video player renders the output to.

And if you want to see what would happen if you treated pixels like individual square polygons, then just do a 4x resize with nearest neighbor and halve the resolution with bilinear.