Is it possible, in this day and age of machine learning, to upscale pictures without loosing quality yet?

Is it possible, in this day and age of machine learning, to upscale pictures without loosing quality yet?

Attached: mpv-shot0001.jpg (1280x720, 113K)

Other urls found in this thread:

waifu2x.udp.jp/
is.tuebingen.mpg.de/de/news/from-small-to-not-so-pixel-perfect-large
twitter.com/AnonBabble

yes

how and where?

svg :)

fan arts

Attached: big-316643638b.jpg (1920x1200, 1.09M)

Looks like shit compared to the original. Pretty good effort though.

Attached: aienmathehellgirlanime1182970_600_447-001_515571_37431.jpg (600x447, 35K)

You can upscale images without loosing quality. You can't gain additional quality though.

This is your best bet:
waifu2x.udp.jp/

Get a proper frame from the original, not an already upscaled mpv shot.

Good post. Good digits.

Attached: waifu_upscale.png (1200x894, 942K)

Surprisingly, we still can't accurately generate visual detail out of nothing.

Was it ever even possible to *lose* quality while *upscaling* pictures anyway?

Actually yes. A poor upscaling algorithm will just give you copious amounts of blur that won't go away once you scale the image back to its original size.

Some scaling kernels blur images if you use 1x "upscale". These kernels make your image lose quality for slight upscales (like 1.1x).

>scallin kernels

Attached: GVJ06OK.jpg (960x540, 111K)

Not so long ago I heard about this:
is.tuebingen.mpg.de/de/news/from-small-to-not-so-pixel-perfect-large
So it is possible, without losing perceivable quality, but of course the algorithm can't replicate data,that wasn't there in the first place, so it is not pixel perfect. It's more like an artist, that fills the blank spots

Use FSRCNNX shader in mpv.

Try nearest neighbor.

Paint?

Attached: upscaled.jpg (6750x5028, 3.33M)

As Sup Forums can't upload more than 4mb
here's a printscreen/screenshot

Attached: printscreen of 12mb.png (1451x1080, 2.52M)

>Is it possible, in this day and age of machine learning, to upscale pictures without loosing quality yet?
Can't even get the spelling right, so no.

print the image with high quality photo paper and then scan it back onto the computer

waifu2x

waifu2x-caffe just works.

Yes, in action movies since 2000

ENHANCE!

Does this actually work

not using the current raster formats. vector formats work well for scaling but grow unacceptably large when trying to store anything with a lot of small details.

the problem is with how scaling works. scaling down is easy because you just average the colors of the pixels. a blue pixel next to a green pixel will become a single cyan pixel when scaled down. when scaling up you can't reverse the averages to find the original values that went into it to begin with because image formats don't store any information in that regard. the best you can do is make guesses, and guessing leads to quality loss.

of course not

Thank you based user
Too bad my monitor is UHD

I always thought in these movies those cameras were already capturing things at 8k+ res, and 'enhancing' was just zooming in.

that's absurd

all you have to do is hit the + sign and say enhance enhance enhance

>Not being like Blade Runner and seeing around corners.

It's called vector graphics.

just take pic, go to paint, and resize percentage to 300%, and save as PNG
As the original picture isn't perfect, the resize can only look as good as the original.
There's fine spot where the size is large enough such that when you make it a fullscreen it doesn't get fucked, nor too small that it just expands with out "increasing" pixels.

It works (the Paint way) the same way as when you look at a picture on a monitor with more pixels, and then another monitor with less pixels, the picture doesn't magically become worse cause it's on a new monitor even though it's technically upscaled. (If you can comprehend what I just said)

Attached: upscaled 3.jpg (6576x4120, 2.49M)

They always made it look like they could zoom-in into a shitty pixelated portion of an image and then make it look as detailed as the original picture

Which is imposible

Attached: 1462321038466.jpg (543x579, 25K)

>Is it possible, in this day and age of machine learning, to stop misspelling "lose"? Kys illiterate.

>being this dumb
How?

Attached: REEEEEEEEEEE.gif (400x204, 44K)

Most images have information beyond the pixels current algorithms just aren’t advanced enough to really use it to the fullest yet.

waifu2x.udp.jp/ is pretty decent, but don't expect it to do anything you see in the movies.

How are you supposed to get more detail out of nothing?

It's not out of nothing it's from prior knowledge and probability of the detail. Sure it's not guaranteed but if it has a good likelihood to be correct then it's still relevant.

not out of nothing, out of something

I hope you are joking.

>printscreen
just resize the damn picture.
damn Sup Forums is full of retards.

>How are you supposed to get more detail out of nothing?
Interpolation.
You have some ability to guess what "should" be between the pixels.

Suppose you had an all green image with 1 pixel missing, guessing that the one pixel is *probably* green too, isn't that complicated, that basic principle can be applied to increase the resolution of a image in a way that makes it also look "higher resolution".