I may have made a mistake, and i welcome your redo if i have, and will call it a good catch.
I started with a converted-to-greyscale square 45 deg screen pattern in photoshop.
As another poster pointed out, the comparison is between the second and third images relative to each other, not necessarily the source.
I "may" have saved a psd of the original 4k, and will be happy to send it to you if truly interested?
Either way, its a relative representation that started a discussion, and so may have some intrinsic value.
@Tim L - it's confusing because what is referred to as 4K is actually 8K. 1080 is 2K so yes '4K' is four times larger. For some reason it was decided to call it 4K to represent it's longest side rather than it's area.
Well, the number of pixels is 4x, so my result may be off by a factor of 2.
Right or wrong, it is a valid representation of what really goes on with bicubic smoothing.
Anyone care to try with Vegas smart resample?
The closest I can come to counting the aliased steps, the 4k represents a ratio of ~17:5 over the downscale.
Close enough to 4:1 or 400% for purposes of demonstration. Guess I'll let it stand as it is.
(3840x2160) / (1920x1080) = 4/1
All of which is pretty insignificant, because we want to look at the results of downsampling, not downscaling, and make our own judgment whether the edge smoothing is beneficial or detrimental to the sharpness of native 1080p.
You guys are now ready to embrace my "Upscaling HD to 4k gives you MORE colors proofs"! :D (same with 4k to 8k...)
Best comeback ever, @wwjd !
hahaha! It does though, I tested it. :) The software logic fills in the inbetweens. BAM!
Show me?
@Brandigan Yes, I read that when it was first published. David Newman (Cineoform & Gopro CTO) is a reputable source. But note David is only commenting on chroma subsampling nothing else.
@Kinvermark Missed this in the noise. Yes, but chroma subsampling is the point. 😉
The resolution of 4K>1080p is still only 1080p, but you get better colour variation when you downsample.
OK, this is a very simplistic overview, but can be extrapolated to what's happening. Take a monochrome 2x2 grid of pixels on part of a 4k sensor, and they've collected light such that there is 1x black and 3x white pixels.
With a camera that's being as fast as possible at creating a 1080p image from a 4k sensor and not downsampling well internally, but simply scanning every alternate line or column (or fudging it in some other way for speed reasons), it might choose a Black pixel and it might choose a White for the 1 pixel it uses to make the 1080p version of the image, depending on where that black pixel fell.
However, if you are downsampling outside the camera, and combined those pixels (ignore all others for now) the average for that 'single' pixel is 25% black and 75% white. That makes all the difference at edges and in detailed areas.
As a side effect/bonus, if the camera moved by the distance of 1 x 4K sized pixel, then the camera's 1080p pixel might now change from black to white or vice versa, whereas the downsampled 1080p pixel would change by less; even if there were now either 4 white pixels, or 2 black + 2 white ones (depending on what new pixels were being scanned from the target image). It goes from 25% Black to either 0% or 50%.
It's more complicated than that, because sensors aren't monochrome and the Bayer mosaic has twice as many Green "pixels" as Red and Blue; but maybe it'll help let you understand, very approximately, what's happening.
The quality of the result is a product of the difference between the camera taking as close to zero time as possible to do its job and software taking as long as it needs to get the better result.
Fast, Good, Cheap. Pick any two. 😉
Show me?
You can test this yourself in yourself in any NLE, with any HD footage. Render it larger, zoom in and look at it. You can SEE the tween colors with eyes. Or use some fancy color counting software. The inbetween pixels are shades (new colors), thus more color. Is it REAL LIFE accurate? Of course not.... but it COULD BE also. It's no miracle cure for 8-bit color grading, but it can help if needed.
I tested this years back, proved to myself, and moved on. Not sure I kept the pictures, but it's an easy test.
Upscaling 1080p to 4k is generally doing something fairly simple like just creating pixels that are halfway between each 1080p pixel and its immediate neighbours, so there is less aliasing when viewed on a 4k screen. That's not particularly colour accurate, but can give more pleasant results than 'the jaggies' you'd get if it just made each pixel 4x larger.
Where that can help is if you've got a particularly 1080p bad image and you scale it up to 4k, apply filters, reduce noise, colour correction etc. then scale it back down again. You'll get a potentially cleaner - certainly different - result than if you just attacked the original at 1080, but YMMV.
Yes, wwjd, bicubic interpolation is the really fuzzy way Vegas does upsampling. Still the same number colors in the.palette, just some filler pixels whuch are anything but smart. We call it noise.
Following that logic, one should be able to actually reach or exceed the container capabilities, by upscaling, then downscaling, the upscaling, then downscaling, .....
But you bring up an interesting distinction, that between upsampling and downsampling, or interpolation and averaging, if you will.
Because downsampling relies on a regression algorithm, differences between traditional downscaling methods are often immaterial. In fact, the results shown in my first image are identical using bilinear sampling. Identical? Now try that theory in an upsampling test!
In all fairness, my reading of authentic sources indicates that true supersampling would leave room for some intrinsic relative gains in post, depending of course on the kind of hardware downconversion already going on inside the camera, and its accuracy relative to the software model being tested. I concede that point; however:
Unfortunately, Vegas supersampling has never worked that way, at least during the Sony days. Only worked on internally generated source. This suggests that all the downsampling pundits may have merely shown up at the wrong sandlot, and that Vegas may not be the optimal venue for performing this type of demonstration.
https://www.vegascreativesoftware.info/us/forum/supersampling-with-photo-videos--83563/#ca489367
Is that why we are seeing so few visual examples here despite the abundance of encouragement? Is that why folks are having such a hard time displaying what they "already know?" It would be hard to imagine that such dedicated hobbyists could merely be so lazy.
But in fairness, it is not just up to Musicvid to prove his point, others must also prove theirs - otherwise it is just an unsubstantiated claim.
No one has stepped to the plate with examples of Vegas' "smart" resampling, so i guess it is a nonplayer as well. The open source Lanczos resizer is unavailable in Vegas afaik, but my experiments indicste it does a better (sharper!) job in post than bicubic, as is claimed.
Hardware uprez has always enjoyed a reputation for being superior to softawre solutions for.delivery and.playback. To conversely claim that all software downsampling is superior to available hardware solutions is illogical as a bicondtitional, and begs for realtime temporal proof, at the very least.
Either way, the purpose of this thread is not being met, so if more visual tests are not forthcoming, and this rumination continues past the weekend, im going to ask Nick to close the gate on this thread, perhaps something he wishes he'd already done.
IF I HAD TIME (and Interest) I'd use my GH4 and do some tests, but my video plate is really full in 2018. Since I'm swimming in 4k now, I just don't care to bother with HD resolutions anymore.
My point was: "1080HD" cameras that really only put out 600 lines, just look worse than 4k down scaled to FULL 1080 "HD". Thus, yes, 4k downscaled looks better.
No posts that i have read in the past week, including my own, have disagreed or challenged that premise. In fact, i "strongly agree."
IF I HAD TIME (and Interest) I'd use my GH4 and do some tests, but my video plate is really full in 2018.
If you have time and interest to post theories, you have time an interest to post your images. And that is my one and only interest in this swarm of bats and bees, I promise.
You "used" to do quite a nice job of that in the past, as i recall....
I think we have enough evidence to suspect that a big part of the solution lies herein:
Kinvermark wrote on 1/26/2018, 7:11 PM
Not hard to understand, but I think a healthy case by case scepticism would be wise.
3) My personal experience using a Panasonic GH4: the 1080p was not very good, so yes, in that case down-sampling from UHD probably makes sense. In this case I think you are seeing the difference between "good" down-sampling in the computer vs "bad" down-sampling in the camera (it all starts from the same sensor readout so it needs to be down-sampled somewhere). I believe this difference is not apparent on the GH5 (yet to test.)
This explanation, of course, all falls apart if you are comparing different cameras. e.g. Canon c100 HD vs GH4 UHD. - I guess this is the point of view that Musicvid and David-Tu are coming from. Also, some cameras can "crop in" to get 1080 from the sensor without binning - so this would be a very clean signal that may be better than the whwho sensor down-sampled.
But since visual feedback was so scarce this time around (here's how our surveys used to be, guys):
Including the first thread, we got three or four replies with evidence out of over a dozen 4k cameras owned. Direct requests for examples from these owners resulted in at least 15 responses with excuses, indifference, or bluster with no examples. Whether due to a shift in culture at the message board level, or something else, i know one thing for certain: If it smells like a kettle of fish, there's probably one nearby.
Cameras owned by responders:
Gh5 (2)
Gh4 (2)
Xc15
A7S
FS7
XA20
iPhone8
F2300
y1 action cam
PXW-X160 HD
RX10 iii
One or more GoPro
permission to lock, Captain.
I'm at my day job posting. ;) No camera or fun video time here. I think I'm confused on the question and test: film 4k and downsample it, also film HD (even if the camera doesn't record a full HD 1080 res) and compare, right?
Did a test TOO SLOPPY for most to accept - I moved camera forward to reframe and compensate for the crop diff tween GH4 4k and HD, but it wasn't enough. Still, HD looks like mush compared to 4k downscaled.
How would you set it up so the framing is the same?
An objectified comparison seems worthwhile, and yet it's still a good effort.
I got this in a fortune cookie from a mediocre takeout today, four hours ago to be exact:
He who has imagination but no learning, has wings but no feet.
Sadly, I have not as much imagination.
musicvid, I should have pushed camera forward a lot more so it was the same size. instead, see the out of focus at the bottom? That is my window, and I used THAT to align, rather than that house down the street.
Yes, wwjd, bicubic interpolation is the really fuzzy way Vegas does upsampling. Still the same number colors in the.palette, just some filler pixels whuch are anything but smart. We call it noise.
Although this thread is moribund; just to clarify to anyone else interested:
Yes, there are still the same 16.7+ Millions colours available in the 8 bit colour pallete. The process of upscaling means that extra pixels are generated - interpolated - from this pallete, between the original pixels. I.e. more of the available colours are used in the final image for a more pleasing result.
Easily verified by a Black box on a White background upscaled from one resolution to a higher one. Unless the program used is 'dumb' and simply stretches pixels (avoid such a product) then grey pixels will be generated at the boundary between Black and White. This may be difficult to see - or always required - on vertical or horizontal lines, it is clearly evident on diagonals, circles etc. which will appear smoother and without 'jaggies'.
No, its not "smart"; just mathematically accurate. I'm not sure "we" call it 'noise'. "We" call it what it is: upscaling by interpolation. Noise is random, this process is not.
The noise in Vegas is not random, as it uses patterns for both downscaling and upscaling. Agreed.
Not necessarily the case with other engines, which I've also tested. So now its become "pseudo-randomized adaptive interpolation?" Think I'll stick with "noise."
Now, you are still working on your own tests, Brandigan? I was hoping you would be able to help wwjd with his camera framing, since moving it destroys objectification, as you so eloquently pointed out earlier in the discussion. C'mon guys, you speak the same language, put your heads together. Got charts, anyone? Remember, once we've agreed on some snapshots, temporal testing comes next (I may be able to help with that).
Actually, this whole thread is just a ploy to aid me in choosing a camera for wildlife photography. Thanks for all that's been offered!😀
Wasn't the object to prove 4k downscaled looked better than ALMOST all HD normally? Reframed result would be the same since GH4 HD is not "full" HD either, but closer than Canon's prosumer HD. Either way, HD always looks mushier (in Canon fan-boi terms: "softer") than 4k. I think I have some charts laying around - see if I can do those during some shoots.
What cameras you looking at Musicvid? They are all GREAT these days, just different flavors and parts. :) There's no reason to get less than 4k these days.
The noise in Vegas is not random, as it uses patterns for both downscaling and upscaling. Agreed.
Not necessarily the case with other engines, which I've also tested. So now its become "pseudo-randomized adaptive interpolation?" Think I'll stick with "noise."
Now, you are still working on your own tests, Brandigan?
"Other Engines"? Is this like '"We" call it noise' and your previous inexplicable claim about "dithering"?
Are you conflating the downsampling process with upscaling? Umm...no, they are quite distinct.
Now, Musicvid, have you got any proof? Not that it's relevant to the original discussion anyway; especially if people are using... Vegas.
If it's doing it correctly - and we have no reason to suppose it isn't without that proof (no rush) - Vegas does not use "patterns" for 'downscaling' [sic], or downsampling either. As you previously seemed doubtful about explanations of the method used in the downsampling process: I'd be interested in what evidence you might now have to support your assertion.
My own tests with downsampling were conducted years ago and I now just do it. I won't be providing anything other than commentary and correction. 😉 You wondered why no one was contributing to this warm, friendly thread? I suspect a bit less of the condescension and passive-aggressive: "Ooh, are everyone's cameras broken?" vibe might have got more people to provide results before, but it's pretty unlikely now. 😆
Did a test TOO SLOPPY for most to accept - I moved camera forward to reframe and compensate for the crop diff tween GH4 4k and HD, but it wasn't enough. Still, HD looks like mush compared to 4k downscaled.
I wouldn't worry about it. You've proved to your own satisfaction that downsampling works.