Category
🤖
TechTranscript
00:00What's up guys, MKBHD here and one of my favorite things about making videos with this high
00:11end gear is the images that these cameras can capture is legit absolutely incredible.
00:17Behind the scenes what you might not know is the actual video file that I'm uploading
00:20to YouTube looks significantly better on my computer than on YouTube.
00:27Like it looks pretty good on YouTube and you can watch in 4K and it looks pretty crispy
00:31and sharp and but it looks miles better when I'm looking at it in QuickTime on my computer
00:36after I've exported it to the point where sometimes I'm like why do I spend so much
00:40time on like the little things just little color adjustments little tiny sharpening and
00:45stabilization and things like that when at the end of the day you're just gonna watch
00:48it on your phone or on some laptop and not really notice 99% of that stuff and it's of
00:54course because that process is so fun but it's also become kind of a running joke that
00:58like if you want to shoot at a crazy high ISO or if you have a ton of noise you don't
01:02really need a denoiser because YouTube processing will just clean up all that noise for you.
01:07It'll just soften it up.
01:08Anyway all this is because YouTube has to compress the file we upload.
01:14So in order to process all the millions of hours of video being uploaded to the site
01:18every month and process them into a 4K version and a 1440p version, a 1080, a 720, a 480,
01:25all these versions and dynamically switch between them all, they've got to do some smart
01:28streamlining and part of that is crunching down the video files to as small as they reasonably
01:34can.
01:35So if I upload this video for example, this beautiful crispy probably 5 gigabyte or so
01:40final file when it's all done and rendered out, YouTube isn't turning around and playing
01:45back that 5 gigabyte monstrous file to you.
01:48Your internet's not good enough to do that in the first place.
01:51What they've correctly assumed is that they can compress it all the way down to just a
01:55few megabytes and it'll look pretty much just as good and it'll speed things up significantly.
02:01So that being said, I've always been curious how serious is this compression?
02:08How serious is YouTube processing in 2019?
02:14And it turns out sort of the best way to test that would be to upload a video and then
02:19download that video and then check out the difference.
02:21But of course, one time's not enough.
02:23You got to do it, I guess, Mr. B style basically at this point.
02:25So the plan is upload a video to YouTube, let it process, then download that video,
02:32then upload that video, let it process, download that one, upload that, download it again and
02:37keep going and keep going and keep going until it breaks further and further down and we'll
02:43do that a thousand times.
02:46So this idea, this has actually been done before back by a channel in 2010 about a decade
02:50ago called Ontologist and they did this with a webcam video, but I suspect with changes
02:55in scale and different codecs, YouTube's processing has evolved a bit since then.
03:00Plus they now support HD videos.
03:03So here we are at the end of 2019.
03:05Let's see if YouTube's processing has gotten any better or any worse.
03:10Let's see how good it really is in 2019.
03:17All right, so I'm sitting in a room different from the one you're in right now and I'm recording
03:22both the sound of my voice and an image of myself and I'm going to upload it to YouTube,
03:29rip it from YouTube and then upload it again and download and upload again and again and
03:34again until all the original characteristics of my voice and my image are destroyed.
03:42This is a test of YouTube processing in 2019.
03:48Okay, so this is the original file.
03:50It's 8K, it's super sharp, super clean, white background and you know, it's me moving around
03:56a little bit.
03:57High quality audio as well.
03:59So this is sort of your classic talking head tech video.
04:04This is our base, our starting point.
04:10All right, so I'm sitting in a room different from the one you're in right now and I'm recording
04:15both the sound of my voice and an image of myself.
04:19So this is file number two.
04:20So this is downloaded from YouTube's native downloader.
04:23So in the YouTube video manager, they give you a download option and this gives us back
04:27a 720p version.
04:29So it's going to be 720p from here on out, which kind of sucks and I almost sort of wonder
04:33what it would have been like if we'd used a third party one that would give us a 1080p
04:38version or something like that, but those are not as consistent.
04:40So just using straight up YouTube downloader and looking at the file, you know, of course
04:44there's less sharpness, but it generally basically resembles the original file, just a 720p version.
04:50Nothing too crazy.
04:55All right, so I'm sitting in a room different from the one you're in right now and I'm recording
05:00both the sound of my voice and an image of myself.
05:03All right, so five downloads in, it's getting softer here for sure.
05:08And we're starting to see the effects of not just losing detail, but the compression algorithms
05:13at work that are trying to save detail and be efficient.
05:17So basically a video compression algorithm will take a look at this entire video, it'll
05:20scan the whole thing and essentially divide it into a bunch of smaller blocks or groups
05:25of pixels.
05:26Then it'll look through the video frame by frame, one by one by one.
05:30And if a certain block of pixels remains the same through multiple frames, it'll know that
05:36it doesn't have to change much.
05:37It can keep that original frame through and save some space.
05:40This whole process is called block motion estimation.
05:43It's very common with video compression and I'll link a really good video below that explains
05:46it super well.
05:47It's only got a couple thousand views, but I think it does the best job.
05:51But basically the way we notice this is certain parts of the video that don't move, like the
05:55background, for example, they will stay fine.
05:58But the parts of the video that do move, like me, it'll start to break down a little bit.
06:04All right.
06:07So I'm sitting in a room different from the one you're in right now, and I'm recording
06:12both the sound of my voice and an image of myself.
06:16So here we go.
06:17Now we're getting some more dramatic compression.
06:20It's super soft overall.
06:21Of course, my face, the details in it.
06:23The skin is just super soft.
06:25But again, the background, which was already soft, isn't losing much detail at all.
06:29In fact, it kind of looks the same as the beginning because it isn't moving from frame
06:32to frame.
06:33They're just keeping those old pixels until my arm passes over in front of it or something.
06:39Then you can kind of see it reset for a couple of frames.
06:41Also by now, you've probably noticed the audio is a bit different.
06:45Turns out YouTube sort of shifts the audio or cuts back the audio by about two frames
06:52with every download and upload, which isn't really enough to notice at first, but now
06:56it's adding up for us to be able to see that difference.
07:01All right.
07:04So I'm sitting in a room different from the one you're in right now, and I'm recording
07:09both the sound of my voice and an image of myself.
07:12Okay.
07:13So we're now we're 20 downloads and 20 uploads in.
07:16And we've gotten to the point where you can, you can kind of start to see the blocks or
07:22like the bigger chunks of pixels that the compression algorithm is sort of holding together.
07:27It's kind of like a math problem.
07:28Like if you just start rounding 15,555, which is a pretty precise measurement, if you start
07:33running that digit by digit over and over and over again, eventually you just get 20,000,
07:40which does that make sense?
07:41But what's funny is you still notice there are still moments when like I also pause,
07:46I'll stop moving for a couple of frames and it'll also sharpen up again like that right
07:50there at 27 seconds for a second there, there were consecutive frames that I didn't really
07:55change too much.
07:57And so it duplicated the pixels and it looked clear for a second.
08:00And then I move again and it goes right back into chaos.
08:03All right.
08:09So I'm sitting in a room different from the one you're in right now, and I'm recording
08:13both the sound of my voice and an image of myself.
08:17Okay.
08:18So the audio shift continues.
08:20This time the audio didn't start for a full two seconds, two and a half seconds into the
08:23video.
08:25And the first instance of what I'll call brutal blockiness is also what's happening.
08:30It's grouping like blocks of blocks.
08:32So like bigger sections of pixels because they're sort of close enough in color that
08:36they might as well be the same color from frame to frame.
08:39And also my skin tone is shifting from brown to more magenta too, if you've noticed that.
08:44And the pauses where I stopped moving for a second are even more dramatic.
08:48Like the whole video frame kind of resets.
08:51So I'll wave my hand in front of the background, in front of the Cheerios for a second, and
08:54it kind of smears magenta all over it.
08:56But when it gets back to a couple still frames, it all snaps back and the smears disappear.
09:01It's fascinating.
09:04All right.
09:09So I'm sitting in a room different from the one you're in right now, it kind of just looks
09:14like now like a struggling graphics card, like someone turned RTX off.
09:18The audio takes four seconds this time to come in and you can now actually kind of hear
09:23a little bit of the audio compression start to take effect.
09:26It's not nearly as drastic as the video, but it kind of sounds like behind a thin cloth,
09:32a little muffled, but not too bad.
09:34Oh, and my waving arms now are leaving disappearing, faint magenta trails on my shirt and on the
09:41wall behind me.
09:43All right.
09:46So I'm sitting in a room different from the one you're in right now, and I'm recording
09:50both the sound of my voice and an image of myself.
09:54It's kind of funny.
09:55The first frame of this one looks like it's going to be okay, and then it all just goes
09:59to chaos.
10:00Now I really look like the faceless cartoon avatar that I used to have on here.
10:04The details in my skin, not only is the sharpness gone, but like my eyes, nose and mouth are
10:09like almost completely gone too.
10:10The magenta trails are getting even more magenta.
10:13They're closer to pink.
10:14Oh, and the sound too.
10:15It doesn't even sound necessarily that much worse, but it's shifted so far now that the
10:21entire last sentence of the video was cut off until all the original characteristics
10:26of my voice and my image are destroyed.
10:29The pauses are so dramatic now that when it finds consecutive frames that haven't been
10:34destroyed yet, like right now it's still around 26 seconds.
10:38And I go from the faceless character to a human and then slowly descends back into a
10:43caricature again.
10:44It's really simplifying the colors here with all the larger blocks of all the rounding
10:49going on.
10:50The background though, honestly, still fine.
10:54All right, so I'm sitting in a room different from the one you're in right now, and I'm
11:02recording both the sound of my voice and an image of myself.
11:08And I'm going to upload it to YouTube.
11:10All right, welcome to my personal hell.
11:13I don't think it can get much worse than this.
11:18Facial features, all the structure, all the contrast, all the detail, all of that is gone.
11:24It's basically turned into pink blocks.
11:26And then the best part now is just like how pristine the background still kind of is.
11:31It's very similar to the original upload.
11:33Even though my actual caricature has turned into a pink, blobbering blob, whatever, those
11:40paintings out of focus behind me kind of look very similar to what they first looked like
11:45when I first downloaded it.
11:53Okay, so this grand finale, it's got it all.
11:56It's part art, part chaos.
11:59Parts of it are completely unintelligible.
12:01You can barely tell it's like just a human being.
12:03That's all you really know about what you're watching.
12:06And then the blocks are even bigger.
12:08Look at my hairline.
12:09It looks like if you turn the graphics all the way down as low as they could possibly
12:13get, then you might end up with that sort of level of Minecraft quality.
12:18And the audio is completely gone.
12:19So it slid from slightly time-delayed to a little bit off to halfway through the video
12:25to sliding completely out the other side.
12:27So it's completely empty.
12:28There's no audio in the thousandth upload.
12:30Turns out the 800th video is about the last time you can hear me say anything compressed
12:35over 800 times.
12:36All right, so I'm sitting in a room.
12:39I bet if we did another thousand downloads, it would look kind of similar to the 1000th
12:43one.
12:44Also, I love that it still triggers the HD badge because it is technically still the
12:49720p file.
12:50But now that this is done, you can see any or all of the thousand uploads on the channel
12:55linked below.
12:56You can also see the auto-generated thumbnails get progressively worse, more nightmarish
13:00over time.
13:01And I'll basically toss in all the important links below to the milestone videos so you
13:04can check them out for yourself.
13:06All right, so I'm sitting in a room different from the one you're in right now, and I'm
13:10recording both the sound of my voice and an image of myself, and I'm going to upload it
13:16to YouTube, rip it from YouTube, and then upload it again and download and upload again
13:23and again and again until all the original characteristics of my voice and my image are
13:29destroyed.
13:30This is a test of YouTube processing in 2019.
13:37So what have we learned from doing all of this?
13:41Well, number one, I will say I think I have more respect for YouTube's processing and
13:46all their compression than I did before.
13:49I obviously still think it looks much better on my computer than on YouTube, but it looks
13:54much better than the 2010 version.
13:56I can say that for sure.
13:57But then number two, the audio held up pretty well aside from the whole shifting issue.
14:01So the actual quality was pretty decent.
14:04It got a little muffled, but it didn't turn into a puddle, which is, I say that because
14:09in the original back in 2010 with that webcam video, have a look.
14:14This is his 1,000th video.
14:21That is hilarious.
14:22So I found this all fascinating.
14:25Was it a gigantic waste of time?
14:27No.
14:28I think someone had to do it.
14:30So either way, that's been it.
14:31I'm happy to finally share this with you after our weeks and weeks of uploading and downloading.
14:37And shout out to the original, which was an audio experiment.
14:41So I'll link the info to that below as well.
14:43With all the rest of the stuff, it'll all be below the like button if you want to check
14:46it out.
14:47And of course, feel free to share this video with anyone who might be interested in a high
14:50quality, super scientific analysis of YouTube processing in 2019.
14:55Thanks for watching.
14:56Catch you guys in the next one.
14:58Peace.