Spears & Munsil 4K UHD HDR Benchmark (Test Discs) May 15, 2023

[Stacey Spears]

Originally Posted by p5browne

Same here on my home desktop PC. Main picture, but placements only for the other 2 pics.

Are you saying that even on your home PC, you don't see the other two? Let me post them in their own posts each and see if that helps.

[Stacey Spears]

Here is the 2nd image

[Stacey Spears]

Here is the 3rd image

[Stacey Spears]

For those that only see the first image. For the images on scaling and dither posts, are all of those images showing up?

If the last two individual posts don't show the images, please let me know and I will have to move them to dropbox and try again.

[wxman2003]

Originally Posted by Stacey Spears

For those that only see the first image. For the images on scaling and dither posts, are all of those images showing up?

If the last two individual posts don't show the images, please let me know and I will have to move them to dropbox and try again.

Still not seeing the 2nd and 3rd images from your past 2 posts.

[gkolb]

Originally Posted by Stacey Spears

For those that only see the first image. For the images on scaling and dither posts, are all of those images showing up?

If the last two individual posts don't show the images, please let me know and I will have to move them to dropbox and try again.

Yes, I see all 8 images in the scaling examples.

[Stacey Spears]

Trying again from Dropbox.

One


Two


Three


And if the above are not showing up, here is a zip file with all three. Maybe someone else can host.

[Stacey Spears]

Sorry this post is so long. What turned out just to be some quick notes on the three stop light images turned into a BTS on the montage grade and some background on our obsession with quality.

The first image used RED Color 2. The second RED Color 3 and the 3rd DragonColor.

RED would release a new color with each camera. It was basically a pre-basked look for 709 color. All of that went away when IPP2 (image processing pipeline version 2) came along. Now 709, 2020 and P3 are the color spaces offered as preset options. Then you can apply your own LUT/look on top of them. Then of course there is a really wide color space space called REDWideGamutRGB or RWG for short. I used RWG on the montage along with Dolby for the color grade. In fact, I use this approach for all of the shot footage (vs. computer generated). For those that want to the technical details on RWG/Lo3G10, here is the paper with the math. Here is another paper with an IPP2 pipeline example. My workflow was:

Use an app called Redcine-X Pro to export a RWG/Log3G10 16-bit TIFF. I would run the TIFFs through our tools to perform temporal noise reduction and deconvolution and output RWG/Linear EXR (16-bit half float). This was sent to Dolby. This is what the ungraded vs. graded will use for the ungraded version. (converted to 2020/2084 just by using the transforms, nothing else creative-wide done)

Dolby used Resolve into the Pulsar as the primary display. A PRM using an eCMU box and an LG C9 using a tunneling box. The LG and PRM were used for tweaking the trim pass settings AFTER the grade was done on the Pulsar. The PRM was used to create the 100 and 600 nit trim passes. Then verified on the C9 to show a consumer displayed with DV worked against the PRM. Any subtle differences were evened out using the trims. No actual primary grade choices were made/changed because of the C9.

One key point I want to make is that no LUTs were used for the grade. LUTs are incredibly popular and used by many, but a LUT has finite precision. Nothing beats a real mathematical transforms and hand grading. The problem is speed. The turn around time using a LUT is much faster and they can pre-create a look and apply it to everything and save so much time. Not to mention there are people like Phil Holland who have spent countless hours building amazing LUT packs based on their years of experience, that allows a novice produce YouTube content that rivals a feature film.

After I left the grading session, Dolby filled the room with every consumer display (at least a sampling of various makes and models) to QC the content before sending it to me. They know the audience for the disc is the enthusiast who likes to pixel peep.

Dolby provided me with 2020/linear (32-bit float) EXRs with the grade baked in. I ran it through Transkoder to export a 2020/2084 16-bit TIFF. Transkoder can deal with the negative pixel values in the EXR to ensure nothing extends beyond 2020 and no hard clipping occurs. Resolve, as of today, hard clips the extreme colors, which is why we have the linear EXRs pass through Transkoder. It does not clip, it correctly remaps.

Then I have one path for Dolby Vision. I scale the TIFFs to UHD resolution TIFFs (the master is 7680x4320). Run Dolby's tools to generate a base layer (10-bit 4:2:0 YUV) and encode using x265. Then I run the next Dolby tool that creates the FEL and MEL base layers and encode those using x265.

For DV MEL, HDR10, HDR10+ and SL-HDR2, I start with the 7680x4320 TIFFs and scale them directly to 10-bit 4:2:0 UHD YUV files using our own tool.

The goal is for each version to have the best starting point. If Dolby would allow me to start with a 12-bit YUV, we would use our dither to get there.

Bottom line, I don't want to use the Dolby base layer for MEL, HDR10, HDR10+ and SL-HDR2 because I don't know how they get from 16-bit down to 10-bit 4:2:0. FEL deals with the extra bits, so even if they rounded for the base layer, the FEL would make up for it.

This is something we will write / draw up and discuss in the future.

You will notice I mention MEL twice, in both Dolby workflow as well as our conversion to YUV. The MEL YUV file is just 0s. It has to be created using Dolby's tools to get the metadata, but we use our dithered YUV file for the video itself. This was not the case on the current disc. Both MEL and FEL are purely the full Dolby approach.

It is a long drawn out over complicated approach,but our goal is to make sure each version is getting the best possible source to start with. It would be easy enough just to use the DV FEL base layer for all versions, but if it had more banding in it because it was rounded to 10-bit vs dithered, it would bother me. The whole point of a benchmark, for us, is to set high standards. We are far from perfect and we continue to make improvements and fixes as we go. Every week I learn something new with Dolby Vision and HDR10.

On the encode side, we set the base layer, for FEL, to 79 Mbps avg and 85 Mbps peak. The FEL is 9 Mbps avg. and 15 Mbps peak. DV on the current disc was encoded by Pixelogic using Sony's encoder. It was the only choice at the time. David, who did the authoring, did some A/B comparisons of its base layer encode an shared with me, our all other montage encodes used x265 and through pixel peeping you can see improvements with x265. We are talking subtle details zoomed in a couple hundred percent. Again, benchmark, so those subtle details matter to us. You can't see those differences sitting in your normal viewing position. The other issue with the Sony encoder, at that time, is that it came in a few Mbps under our target. Maybe a similar issue to what I was having by not using the same settings for both encode passes. Going back to HD DVD, for VC-1, we also did a faster first pass. To give you an idea of time, it is 1.5 hours vs. 30 hours between a fast and slow first pass. :-) All just to make sure I get that extra 3-4 Mbps that I specified to the encoder. This is just an 8-minute clip. Nothing we do would even be remotely practical for an actual movie.

The MEL, HDR10, HDR10+ and SL-HDR2 is 89 Mbps avg and 99 Mbps peak. Actually, I keep tweaking it. We had issues on the last disc where the peak of some versions is lower (~1 Mbps) because they would not mux. I am hoping this solves that problem.

We are now using the same settings (encoder) for both first and second passes. What we discovered was if we used a faster first pass, then the second pass would come in 3-4 Mbps lower than the target. If we use the same settings, then it is within 500 kbps.

The montage takes ~4 days to encode two passes of the base layer. It has been re-encoded many times. e.g. When we received the clipped sun replacement. As we keep swapping out the title card and new end credit cards. Each time tweaking the encoder settings using the latest daily build of x265. Short answer, we are just insane! :-)

We had called our tool xScaler for extreme scaler. It was a cost no object scaler where cost was time. That is one of our fundamentals. Believe me, some things take way too long and I am not patient.

This weekend we just updated the end credit card with the three new people creating the Atmos mix. That means re-doing it all over again.

Next I need to re-render a 1000 nit 2020 and P3 Dolby Vision derived to pre-encode. The 1000 nit version is also used for the SDR vs. HDR compare and the HLG version. This time we are making two HDR vs. SDR. One at 1000 vs. 100 and one at 1000 vs. 203. One will probably be bonus on the SDR disc. Since I personally calibrate my SDR to 100, that is what I consider correct. But, for those that argue no one watches SDR at 100, the 203 is for them.

The ungraded vs. graded was a real surprise for me. I had not realized how much the footage was pushed / stretched during the grade. It almost looks like SDR vs. HDR as well. Unlike the SDR vs. HDR were we limit the peak of HDR to 1000, this version has no peak limit. Goes to show how much usable dynamic range is in the captured footage and the benefits of RAW. Wait till you see the noise / gray in all the shots that have black backgrounds.

Shane even created a special key just to remove sensor noise from those specific shots. Shane really deserves a lot of praise for dealing with my OCD. :-) I can never thank him enough. Thank you Shane, 3000! And Thank you to Dolby for allowing Shane the time he has put into this project over the past year and a half. The credits page on the disc will be a lot longer and more specific this time around. While the names on the disc is Don and my myself, there is no way would could have made it by ourselves.

[Stacey Spears]

Does anyone track the peak audio bitrates for movies on UHD BD with Atmos? Trying to figure out what I need to limit the peak video bitrate to knowing that I will be using an Atmos track. The completed audio is a few months out, but I need to do some video encoding now.

[p5browne]

Originally Posted by Stacey Spears

Trying again from Dropbox.

One


Two


Three


And if the above are not showing up, here is a zip file with all three. Maybe someone else can host.

First picture shows a splotchy Red. 2nd solid Red, 3rd some white mixed in, but at least I see 3 pictures now.

[Geoff D]

Stacey: how does the FEL work if you're not using the base layer for it that you ran the FEL from? Or are you saying that it's the FEL DV encode that uses that specific Dolby-derived base layer and everything else, even the MEL version, will switch to your own personal base/HDR10 layer?

[Wendell R. Breland]

Originally Posted by Stacey Spears

Does anyone track the peak audio bitrates for movies on UHD BD with Atmos? Trying to figure out what I need to limit the peak video bitrate to knowing that I will be using an Atmos track. The completed audio is a few months out, but I need to do some video encoding now.

Only do spot checks using a Oppo UDP-203, many will show around 4 to 5 Mbps with peaks up to 12 to 15 Mbps for Dolby Atmos. I will see if I can find more definitive data.

[Stacey Spears]

Originally Posted by Geoff D

Stacey: how does the FEL work if you're not using the base layer for it that you ran the FEL from? Or are you saying that it's the FEL DV encode that uses that specific Dolby-derived base layer and everything else, even the MEL version, will switch to your own personal base/HDR10 layer?

The FEL is created by decoding the encoded base layer HEVC and then comparing it against the original 16-bit TIFFs the YUV was made from. It then stores that difference between the two in the FEL YUV.

Dither in my YUV file would make it more different from the original TIFFs than a simple round (assuming that is what they do, I don't really know) YUV from the TIFFs and therefore have more difference data in the FEL, making it more difficult to compress.

[Stacey Spears]

Originally Posted by p5browne

First picture shows a splotchy Red. 2nd solid Red, 3rd some white mixed in, but at least I see 3 pictures now.

This was to address the post about including stop lights in the montage. The three images were the exact same image, just different developments (converting the RAW image into RGB). Point being that there are many reasons a stop light or tail light may look different on screen than what you see in real life.

This was an example of differences in the same image with just an advancement in color processing over the years.

I would say the first image saturated the red too much. The 2nd looks almost like the reds are clipping and the 3rd was something more in between. These were done back in 2014, It may look a lot different today as color processing in various apps has continued to improve over the years.

[Geoff D]

Originally Posted by Stacey Spears

The FEL is created by decoding the encoded base layer HEVC and then comparing it against the original 16-bit TIFFs the YUV was made from. It then stores that difference between the two in the FEL YUV.

Dither in my YUV file would make it more different from the original TIFFs than a simple round (assuming that is what they do, I don't really know) YUV from the TIFFs and therefore have more difference data in the FEL, making it more difficult to compress.

Yeah I get how FEL works, but I don't understand what you're saying about using DV on the disc. Is the base layer for the FEL its own separate encode to the base/standard layer for the other versions?

[mrtickleuk]

Originally Posted by Stacey Spears

The FEL is created by decoding the encoded base layer HEVC and then comparing it against the original 16-bit TIFFs the YUV was made from. It then stores that difference between the two in the FEL YUV.

I find this very interesting, since I had previously assumed the two layers were "chopped up", from a 12-bit source, into the first 10 bits (in the base layer) and the remaining 2 bits into the full Enhancement layer. But the way you describe it allows for some?/many?/all? of the HEVC encoding artefacts to also be "helped" as well as adding the extra 2 bits. So it's better than I assumed.

How effective is the FEL at removing such artefacts, would you say (I know it's a horribly general question)?

[Stacey Spears]

Originally Posted by Wendell R. Breland

Only do spot checks using a Oppo UDP-203, many will show around 4 to 5 Mbps with peaks up to 12 to 15 Mbps for Dolby Atmos. I will see if I can find more definitive data.

Thank you, we greatly appreciate it!

I am setting the peak of the video to 99.5 Mbps. The max bitrate for BD is 127.9 Mbps. so I am hoping the Atmos does not go over 25 Mbps. I won't have the actual audio until late August and I am encoding the videos now.

Obviously I will re-encode if the audio goes above 30 Mbps. I leave a little space because x265 may go above my peak. It did last time around, but that may be related to different settings for 1st and 2nd pass. I found I need to use the same settings or the 2nd pass comes in 3-4 Mbps under target, which seems like a lot.

[Stacey Spears]

Originally Posted by Geoff D

Yeah I get how FEL works, but I don't understand what you're saying about using DV on the disc. Is the base layer for the FEL its own separate encode to the base/standard layer for the other versions?

Yes.

I am using the max bitrates for each version. The FEL BL is 80 avg, 85 Mbps peak. When combined with the FEL, it comes in around 90 Mbps average and 99.5 Mbps peak.

If I used the same BL, then the MEL, HDR10, 10+, etc... would be 80 avg and 85 Mbps. I am encoding them at 90 avg and 99.5 Mbps.

I am trying to maximize quality for each version and not limit based on one flavor vs. another.

I don't know how Dolby goes from 16-bit to 10-bit (round vs. dither) for the FEL BL. So, for the others, I use our dither for the best quality possible in the other versions.

I hope that makes sense. If not, let me know and I will try and do better.

[Stacey Spears]

Originally Posted by mrtickleuk

I find this very interesting, since I had previously assumed the two layers were "chopped up", from a 12-bit source, into the first 10 bits (in the base layer) and the remaining 2 bits into the full Enhancement layer. But the way you describe it allows for some?/many?/all? of the HEVC encoding artefacts to also be "helped" as well as adding the extra 2 bits. So it's better than I assumed.

How effective is the FEL at removing such artefacts, would you say (I know it's a horribly general question)?

Yes.

We had the same question as you. I know someone who did a test. I have not done it myself. I think they encoded the BL at 1 Mbps and the FEL at 80-90 Mbps, which is an extreme scenario.

The new montage really shows off those YCbCr artifacts that ICtCp resolves on the 10,000 nit version. The 1000 nit versions don't seem to suffer from the artifacts. At least the worst ones, which you can see easily see on the tulips and peacock feather. Sadly BD is YCbCr only, so I will need to make an ICtCp mp4 post launch and find a way to deliver to customers that purchased the disc. Then they can play from a USB drive.

We are looking at some papers, like one from Netflix. The problem is the papers don't really cover the encode side, so not sure how well their idea survives compression quantization.

[p5browne]

For the non DIYers out there.
HDMI settings for your set, are a different set of settings, than from those used by the USB inputs. Just to confuse things more, the underlying settings are the SDR settings. Not playing a video - SDR settings. Playing a video is required, in order for the settings, for what ever parameter you're using. (HDR, DV, USB, HDMI, etc) for installing, changing or reading. (Else wise if changing settings, and no video playing, you'll mess up the SDR settings)