For me, it's Catmull-Rom.
>>100189874for me it's Ur-Mom lmao
this looks like they were scaled without being gamma corrected first
>>100189938Gamma correction is a meme, most applications like image viewers and media players don't even apply gamma correction while scaling.
>>100189874>caring about this in the age of AI upscaling and coloring everything easily
>>100189874If you're using anything but None/Nearest at integer values for upscaling you're doing it wrong.If you're downsampling, well, that's more forgiving.
>>100189965that's nice but the ones i use do
>>100190083Pixels aren't squares. Also you will get uneven pixel sizes/image warping if you nearest-neighbor scale to a noninteger factor, which is a very common issue for emulation. The main reason why all of these pointless resampling algorithms even exist is because nearest-neighbor looks like shit.>>100190114Bullshit. Linearization is too slow for most low-power hardware like shittops and shitphones. There is no application that defaults to gamma-correction for compatibility reasons.
>>100189874For me, none and nearest
>>100190083>if you're not doing [objectively shittiest option] you're doing it wrong
>>100190055>age of AI upscaling and coloring everything Even you know this is bullshit, because if this was true then simple acts such as window scaling would be unbearably slow. Tensor cores aren't magic and will always be leagues slower and more inefficient than fixed-hardware GPU interpolation
>>100190188>Linearization is too slowwhat? no it isn't. if you benchmark it with ffmpeg or mpv you'll see it doesn't even affect frame times because the bottleneck is memory speed unless you're using a meme scaler
>>100190437ffplay doesn't even default to linearized scaling. And neither does mpv.
>>100189874for me it's integer nearest neighbor.It's annoying when i'm making "pixel" art projects and having to prevent browsers from doing something retarded by default. it's insane.
>>100190188>Pixels aren't squaresPixels are pixels and are pixels.Pixels will always be pixels.You can represent them as dots to manifest them into the real world, or with RGB stripes, or with pantone color squares under D50 illuminants for a tiled mosaic.When it comes to RESAMPLING pixels, there's no good way to upsample them. Just enlarge the pixels, and be done with it.If you have a pipeline that treats them as squares (digital displays, typically, despite them being RGB stripes in physical form in the real world) then you just turn one square into 4 squares to zoom in 2x.>Also you will get uneven pixel sizes/image warping if you nearest-neighbor scale to a noninteger factorNo shit.If you must do this bad thing, then it doesn't really matter what you use, the end result will be pretty shit regardless.>The main reason why all of these pointless resampling algorithms even exist is because nearest-neighbor looks like shit.No, they're simply using it wrong.NN+Integer values looks perfect, if you think it looks bad, then that's your content that looks bad and you're misplacing blame. Pixelation ain't such a good look, but it sure beats blur or AI slop.
>>100189874ARR ROOK SAME!!
>>100190055AI upscaling is heavy and looks like shit at anything lower than 1080p>>100190083Nearest is only good in pixel art
bilinearbros ww@all nearest neighbour niggers must fucking hang
>>100189938The pixels could do with being converted from RGB to HSL before being interpolated. The results here look muddy.
They all look retarded, what fucking use case would someone need for those?
For me, it's rendering onto a buffer 1/10th the size the assets were designed for and then up-scaline using Bilenear because I'm too lazy to learn how real console limitations work and I need to jump on the PS1/N64 era nostalgia wave before it's too late.
>none
>>100189874Just use ewa_lanczossharp with AR until the AI scalers catch up in accuracy and power consumption
>>100189874none/nearest looks best.Everything else is a blurry mess.
>>100191308neither the ps1 nor n64 used bilinear interpolation
>>100189965Absolute midwit take. Everything looks darker than nearest without it.
>>100191450trust me bro, this is retro. just add some inappropriate dithering and some high-contrast neon colors like totally all of the PS1 and N64 games had
>>100191450N64 had a filtering that was like bilinear but worse. Both Mupen and Duckstation have options for bilinear texture filtering and in full 3D games it actually looks pretty good
>>100191355noooooooyou heckin have to resamplerino the pixelerinos, they're not actually squares! they can be blurred and stretched to look not pixelated!!!seriouslyjust resizing the pixels is very BAD because you can see pixels! okay?
>>100190253thats how you get yous!! say the stupidest "wrongest" shit ever and receive the dopamine
>>100191467both consoles had dithering, most games output 15bit colour, though you're unlikely to notice dithering with n64 games due to the hardware anti-alias filter which blurs... everything>>100191547the n64 has a rather unique "3 point" bilinear filter, rather than the normal 4 point bilinear, the difference is quite obvious when looking at a texture close up, which you always do on the n64 since the textures were so smallthe psx uses a point filter, or in modern terms, no filter, which is easy to replicate, but psx has its' own challenges in terms of replicating how its' graphics worked, notably the fact textures are affine transformed, no perspective correction, and also the fact the 3D math is all done in fixed-point, since it has no floating point unit, making vertices very "wobbly"
People yearn for the seams.
>>100189874nice artificial test casenow show some comparisons of real-world samples
>>100191636Good thing modern emulators can eliminate the wobbliness of PS1 games.
>>100191670Yeah, so great that it's possible to intentionally make your game look like complete shit.
>>100191547>>100191636it's worth noting that simply using normal bilinear isn't necessarily an upgrade, games were sometimes made to work around the look of 3 point filtering, which you lose if you do that, plus just in general it doesn't look the same, using bilinear just "looks like it's emulated", since that's all emulators could do for a long time>>100191670i was so amazed when i first heard about that, at first i wondered why emulators didn't fix that, then i read into why it happens, so for a long time i understood that it wasn't possible, then they go and find a way anyway>>100191715i think you misread his point, he said "eliminate", not "emulate". until recently emulators reproduced the same wobbly vertices and not because of some desire to be faithful, but because it was an actual technical limitation of the console, all the math was done with integers, so while you could scale the graphics up, vertices were still stuck on the original, very coarse integer grid
>>100191760Overlay is beautiful and crisp on the right. Looks like shit on left.Clever filtering makes the edge of stairs look great.
>>100191670it's crazy what modern psx emulators can dohttps://www.youtube.com/watch?v=bn-Udeo96Lkstable vertices and perspective textures are nothing special on n64 emulators, since the n64 did those on its own, it's not trivial at all to implement them in a psx emulator
I like Lanczos because the painterly effect is cool
>>100192792Painterly effect? You mean the antialiasing?
>>100191451>noooo everything must be linearized because.. because it's incorrect otherwise!!holy midwit. As i already said, your browser doesnt linearize while scaling, youtube's player don't linearize while scaling, your image viewer doesn't linearize while scaling, VLC/ffplay doesn't linearize while scaling. Your attempts to get "correct" results are futile, and no one in the industry cares. Even studios who print 1080p BDs (which are just downscales from 4K masters) aren't linearized. They're all gamma uncorrected.
>>100193523yes
>>100190083This. NN when upscaling, lanczos when downscaling.
>>100193587Cope.
>>100189874I like spline36 and lanczos, altough they are a bit too slow, in which case I'd rather use catmull-rom, then mitchell, then hermite
>>100191585>okay?What are you asking?
>>100189874scale=ewa_lanczossharpdscale=catmull_rom
scale=ewa_lanczossharpdscale=catmull_rom
>>100189874there's a thread on /gif/ where the OP is a webm of a girl shitting out candy and that shit looks like pic related
>>100194889ACCEPT THAT WE MUST BLUR THE PIXELSLUDDITEgoshur so stubern and stupitthe human eye does not see in pixelsif you look at computer screen or television and don't have to ask "what the fuck am I looking at?" you are doing it wrong and need to blur the pixels more until that happens
>scan drawing I made in 1200dpi>decide that its too wide, scale it lower horizontallyIs there any point to using something besides nearest? At that resolution I don't think it's noticeable either way and in the end it will then get scaled down to something that actually fits on a computer screen.I also kind of like the sharp look of scaling down with nearest
>>100196558Did you fail your computer graphics course or something? What's with this unbridled seething at basic low-pass filtering?
>>100196695uneven pixel sizes
>>100196756But I'm scaling it down from an absurdly high resolution anyway
>>100196787you're still going to encounter that issue, and even worse, your image will be full of high contrast and high frequency information, which means bad moire and aliasing.
>>100196787>>100196851m-n for comparison
>>100196695If you're downsampling, use bicubic or something at least.Unless you're downsampling by integers, like taking a 1200 DPI scan of an 8x10 drawing and turning it into a square, you can scale that by integers down.8x1200=960010x1200=12000If you scaled this image down to 2400x2400, you can do 25% width and 20% height, integer scaled with NN or Linear with good results.If you do anything even scaling a 1600x900 image into 1280x720 NN won't do good. Use bicubic or leave it in original size, and leave the future user to resample/display as desired without baking scaling artifacts into the original content.In general, always avoid scaling. It's always bad.Even a small adjustment like changing a 4000x4800 image into 4000x4600 will scale ALL the pixels slightly and cause issues.
