Steve Yedlin, ASC on Resolution

[Mike Finegan]

will that test ever be public?

 

[Kemalettin Sert]

will that test ever be public?

afaik he already showed these on camerimage in last 2 years.

 

[joseph slomka]

As always Steve Yedlin's dedication and rigure is impressive..

if anyone close to the test could leak what compression ratios the Weapon was shot at? the wide shot suggested to me it was greater than 5to1!

The compression used was the lowest for the camera

 

[Michael Lindsay]

The compression used was the lowest for the camera

thank you Joseph.. that is useful to hear!...

...the differences between modern digital cinema cameras are not particularly significant?... from a Colour Science how do you feel?.. if you prefer not to answer this publically feel free not to ;-)

this position I feel is very nearly true and is suported by this test...further qualified and refined as- from a 'story POV' within the target of non specialised cinema, TV or internet you will not be overly penalised by picking one cameras best version of raw over the other… very interested to hear if you agree? (from a colour science pov)

Obviously from an a 'ease of use' pov or a hire company biz pov yes there may be a difference! Also cameras assistants, operators, DITs, even grips are effected by the camera selection witout picture quality coming into it.. and even from a brand values POV if people sleep easier there is value in one slection over another !

thanks again

 

[Christoffer Glans]

The compression used was the lowest for the camera

You mean 3:1 or 4:1?

The more I analyse and work with color grading tests the more I feel that because of Reds high noise levels, I'd rather shoot at the lowest compression always for anything. Even a smooth low detail surface benefits from lower compression due to the noise levels being handled better. Especially of you choose ISO800 as the basis for a project. Low compression and a bit of noise reduction in post with some added grain (not too much) creates a very pleasant image.

 

[Marc Wielage]

You mean 3:1 or 4:1? The more I analyse and work with color grading tests the more I feel that because of Reds high noise levels, I'd rather shoot at the lowest compression always for anything.

Michael Cioni has generally said that 5:1 is the sweet spot between quality and economy with Red raw, and I tend to agree with him.

 

[Christoffer Glans]

Michael Cioni has generally said that 5:1 is the sweet spot between quality and economy with Red raw, and I tend to agree with him.

What I meant is that because storage is so cheap today and noise is very much present and gets help from lower compression, 4:1 feels like a better way to get better images before grading, even if it's marginal.

 

[David Battistella]

You know what's curious?

You shoot 5:1 and then add grain to give the images an emotional feel and energy. Because redcod never tends to band, shooting higher compression tends to give you a kind of digital version, specific to RED, of the sensors noise characteristics.

Adding grain is an attempt to connect emotionally with film, but generally that is only going to connect to an audience of a certain age that has that nostalgic memory of film. (If we are talking aesthetics here which is the essence of yedlin's argument .

David

 

[Stacey Spears]

Overall I thought it was well done. Based on his math:

width * height = resolution

7680 * 4320 = 33177600
3840 * 2160 = 8294400

1/4 of them are red: 33177600 / 4 = 8294400
1/4 of them are blue: 33177600 / 4 = 8294400
1/2 of them are green: 33177600 / 2 = 16588800

As you can tell, shooting on an 8K sensor will deliver 4K worth of resolution in red and blue.

I don't get the love for film, especially after seeing this demo. It is so noisy. I was a little disspointed in the 15-perf IMAX performance.

Do we know if the old or new debayer was used? The halation test would either perform better, if the old debayer was used, or perform worse if the new debayer was used.

The one thing I found off was the compression test. I have never seen Blu-ray lose that much detail. In fact, I would say Blu-ray is visually lossless in many cases. At least you don't lose detail in the face like his example. Not even Netflix or Amazon lose that much detail in the face.

I am planning to include 100 Mbps HD, 100 Mbps UHD and 15 Mbps UHD on my next disc, using the same content. (sourced from QUHD) He is right that the quality of scaling will have a big impact on the final results.

Here is the new scaling and deconvolution algorithm that I am using. You can see what it does on my DP Review test clip. Some notes:

1. You can compare the QUHD to QUHD w/ decon only. The decon, or anti-ringing code, is removing the chroma moire in the sweep. I consider this a happy accident. Decon is not a sharpening algorithm. The goal is to recover detail lost by the OLPF. If you go 1:1, you can see the improvement w/o added ringing.
2. I am showing three stages of my scaling, so you can pixel peep and compare:
a. Box filter (area average) This is what I use for a 2:1 downscale. This filter does not ring, but it is soft by itself.
b. Box w/ SSIM. This is the perceptual part of the scaling. Idea is based on this paper. This fixes the softness caused by the box filter. This does add ringing.
c. Box w/ SSIM and AR (Anti-ringing). This is a post anti-ringing filter that reduces the ringing caused by the SSIM process. When scaling up, ringing often looks like a halo. When scaling down, it is diffrent. You often get black outlines.

In the end, I would always use c, but I am including all stages so you can pixel peep. e.g. If you zoom into the black and gray registration markers, you can see the ringing in the black from SSIM. It is not in the box or box w/ SSIM and AR. The goal is not to be super sharp, it is to try and perceptually match the original sized image.

Need to tune the sharpness of the SSIM process. I am planning to tune this using two 32" Dell displays. One 7680x4320 to play the original size. Two a UHD version of the display playing back my scaled version. I will adjust the strength until they visually match. As close as they can. I will need to use several different clips. Also need to tune the anti-ringing strength. It is a bit strong at the moment. It is a trade off on ringing vs. sharpness.

The scaling example is based on the QUHD and not the QUHD_Decon. Did not want that variable in the equation, at least for now.

The point of showing the scaling is to show that the quality of the downscale will impact the final results. I would argue that if he used a better down scaling algorithm, then you would see more of a difference in the final results. Most often, people are using algorithms like Lanczos, Catmull-Rom, etc... All of these algorithms ring and I believe they are the wrong approach to image scaling. I mention using box for 2:1. For >2:1, I would not use a box filter as you can get moire. In the DP Review wall, you do get moire on the wedges. I am looking at using Gaussian for >2:1. I have Lanczos and Catmull-Rom working with the SSIM and SSIM w/ AR, I just don't think it is the best approach. They do not have the moire that box has on the >2:1 (e.g. 7680x4320 -> 1920x1080) but they ring in the kernel itself. I do have the anti-ringing filter for the kernel when using Catmull-Rom and Lanczos, but I am working on Gaussian, which I believe is the better approach. I know that many would think I am crazy for using box or Gaussian. It is possible the Gaussian does not work at all like I hope. A lot of this is you get an idea, try it out and it either works or it does not. We did this a lot when building an encoder.

Here is a nice scaling test of SSIM w/ bicubic as the kernel against all of the popular scaling kernels. In this example, SSIM_AB is probably the best result. (AB is anti-bloat) This was done using madVR. All of your favorite kernels are in here. This is an extreme test image. The source had decon applied as well as sharpening. I recently paid someone on RU to reshoot this using Dragon 6K and Helium 8K so that I can tune based on the Red camera and not a DSLR.

 

[Robino_J]

Adding grain is an attempt to connect emotionally with film, but generally that is only going to connect to an audience of a certain age that has that nostalgic memory of film. (If we are talking aesthetics here which is the essence of yedlin's argument .

David

The youngsters might also find the less plastic'ish look different and interesting. Young people are not restricted or forced to only watch digitally shot content from early 2000s to today, I'm pretty sure everyone have watched celluloid films and might want that look (or emulate it) at one point or another.

 

[Les Hillis]

One side effect of the perfectly clean digital image is that when all character or object movement stops in the frame it can look like the actual program has paused or stopped, especially when viewing content that has to buffer a bit before playing. Adding or shooting for a degree of grain to ensure there's some kind of movement in the frame at all times can counter that disconcerting effect and keep the viewer engaged with the visuals.

 

[David Battistella]

The youngsters might also find the less plastic'ish look different and interesting. Young people are not restricted or forced to only watch digitally shot content from early 2000s to today, I'm pretty sure everyone have watched celluloid films and might want that look (or emulate it) at one point or another.

They seem to love polaroid photographs and vinyl records. And they seems to also like milky blacks, flares and light leaks.

 

[Christoffer Glans]

One side effect of the perfectly clean digital image is that when all character or object movement stops in the frame it can look like the actual program has paused or stopped, especially when viewing content that has to buffer a bit before playing. Adding or shooting for a degree of grain to ensure there's some kind of movement in the frame at all times can counter that disconcerting effect and keep the viewer engaged with the visuals.

If everything stops and it feels like a still image.... you might want to make that edit to move the story forward :tongue_smilie:

 

[Les Hillis]

It only takes a pause for it to happen, but yeah, you could cut around it too.

 

[Marc Wielage]

What I meant is that because storage is so cheap today and noise is very much present and gets help from lower compression, 4:1 feels like a better way to get better images before grading, even if it's marginal.

I have not found that to be the case in all the Red images I've graded. I have seen visual problems once it goes past 10:1, but if I'm asked (which is rare), I'll tell them "don't go to more compression than 5:1."

One side effect of the perfectly clean digital image is that when all character or object movement stops in the frame it can look like the actual program has paused or stopped, especially when viewing content that has to buffer a bit before playing. Adding or shooting for a degree of grain to ensure there's some kind of movement in the frame at all times can counter that disconcerting effect and keep the viewer engaged with the visuals.

We can always add grain in post as a final step. You would be amazed how many TV shows and feature films do add grain as part of the process. I think I must have at least 100 grain samples in my library, all the way from incredibly-bad Super 8mm to extremely smooth ISO 50 35mm negative.

 

[joseph slomka]

It depends on the type of shooting you are doing.
For the most part if you are doing camera tests and properly lighting for the look you are expecting, most high level professional cameras are interchangeable for the bulk of work, even into HDR.
It comes down to operation preference and workflow for the most part.

Basically any camera that can't capture an actor with accurate skintones in a few stops above and a few below will not be on the market for long. In that sense cameras have to be similar.

Steve Yedlins display prep demo show this pretty well. For the most part you can make all of the professional cameras look the same and he does a good job giving a number of different cameras a consistent look. It's exciting from my perspective because post houses do that day in and day out, but can never show the concept in action because we do it for clients on their footage.

thank you Joseph.. that is useful to hear!...

...the differences between modern digital cinema cameras are not particularly significant?... from a Colour Science how do you feel?.. if you prefer not to answer this publically feel free not to ;-)

this position I feel is very nearly true and is suported by this test...further qualified and refined as- from a 'story POV' within the target of non specialised cinema, TV or internet you will not be overly penalised by picking one cameras best version of raw over the other… very interested to hear if you agree? (from a colour science pov)

Obviously from an a 'ease of use' pov or a hire company biz pov yes there may be a difference! Also cameras assistants, operators, DITs, even grips are effected by the camera selection witout picture quality coming into it.. and even from a brand values POV if people sleep easier there is value in one slection over another !

thanks again

 

[Philip grossman]

Graeme,

Was wondering if taking a 6K image, down sampling to 2K then up sampling to 4K is really a good test for comparison to a 2k up sampled to 4K. The reality is that the 6K down sampled image to 2K would still be better then an original image captured at 2K since the original 6K image is over sampled. Not that I disagree with the fact that sometimes resolution is elevated to much, this seems like more of a test of the scaling algorithm since all these cameras start with an incredible amount of resolution.

as always, your insight is greatly appreciated.

 

[Graeme Nattress]

Graeme,

Was wondering if taking a 6K image, down sampling to 2K then up sampling to 4K is really a good test for comparison to a 2k up sampled to 4K. The reality is that the 6K down sampled image to 2K would still be better then an original image captured at 2K since the original 6K image is over sampled. Not that I disagree with the fact that sometimes resolution is elevated to much, this seems like more of a test of the scaling algorithm since all these cameras start with an incredible amount of resolution.

as always, your insight is greatly appreciated.

Starting with a very high resolution source to make the 2k is giving the 2k every advantage. It's making the best 2k possible. In reality, if you started with a real 2k source the situation would be more obvious on how inferior the 2k is.

There's certainly an element of testing scaling algorithms. To me, the best scaling is down, not up, and I don't think you'll get much disagreement there....

Graeme

 

[Michel Hafner]

Here is a nice scaling test of SSIM w/ bicubic as the kernel against all of the popular scaling kernels. In this example, SSIM_AB is probably the best result. (AB is anti-bloat) This was done using madVR. All of your favorite kernels are in here. This is an extreme test image. The source had decon applied as well as sharpening. I recently paid someone on RU to reshoot this using Dragon 6K and Helium 8K so that I can tune based on the Red camera and not a DSLR.

Good to see that these filters are still being researched and new ideas tested. But the differences are now marginal between the best old filters and new ones unless one zooms in quite some.

 

[Joel Arvidsson]

As a metaphor for people that think its pointless to shoot higher then hd/4k, I like to use "shrink plastic" (not sure its the right term in English) As a kid we did alot of them. You paint something on the plastic and when its shrinks in the Owen, and in the process everything gets more smooth and crisp. Your artifact from coloring your artwork disappears. Could not say the same thing when stretching something.