How can I upscale from 1080p SDR to 4K HDR

[Hardback247]

I have the Editor's Cut of Spider-Man 3, which is only available in 1080p SDR, and I would really like to have it in 4K HDR, after seeing just how fantastic the footage of the theatrical cut is in in 4K. I want to upscale all of the Editor’s Cut footage to 4K HDR and splice them into when they appear in the theatrical cut. I don't know what kind of upscaler is best for this, though, and I don't want to use an AI upscaler. Could someone please help me?

[johns0]

Please do not open another thread about upscaling,i deleted your other thread.

[Hardback247]

Please do not open another thread about upscaling,i deleted your other thread.

I wasn't so sure if "Restoration" was the right topic to post about it in.

[ProWo]

I have the Editor's Cut of Spider-Man 3, which is only available in 1080p SDR, and I would really like to have it in 4K HDR, after seeing just how fantastic the footage of the theatrical cut is in in 4K. I want to upscale all of the Editor’s Cut footage to 4K HDR and splice them into when they appear in the theatrical cut. I don't know what kind of upscaler is best for this, though, and I don't want to use an AI upscaler. Could someone please help me?

SDR to HDR10/HLG is not possible because the necessary information is missing (e.g. master display etc.).

[Cornucopia]

Yeah, this falls under the category of "Enhancement", similar to Colorization and 2D->3D, but even harder IMO, since there is not much underlying data on what is lost from HDR version/what is needed to get to HDR version (and also what LUT would be the best inverse equivalent of the Tone Mapped SDR conversion method).

IOW, you don't know what you don't know.


Scott

[usually_quiet]

Even 2160p video upscaled from 1080p or 1080i video can be less than satisfactory compared side-by-side with native 2160p video. The upscaled video can only have the same level of detail as the source and will include some upscaling artifacts too.

[Selur]

just a side note, which is not really helpful for Hardback247:

SDR to HDR10/HLG is not possible because the necessary information is missing (e.g. master display etc.).

not possible is harsh,..
RTX cards can upscale SDR to HDR and send the metadata to my hdr display, .... and .... it's not bad.
(sure true hdr is better, it's still nice)
Only problem is that atm., afaik., there is no software which can capture the generated HDR metadata that which send to the monitor.
=> so don't know a working solution, and what is generated will not be as good as native HDR (same with upscaling to 4k)

[ProWo]

RTX cards can upscale SDR to HDR and send the metadata to my hdr display, .... and .... it's not bad.

And where does the rtx card get the metadata from?
Some kind of average metadata is used, but it is not optimized for the respective film.
So it's just fake HDR.

[Selur]

Sure, it's 'fake' generated stuff by some machine learning stuff.
So if you know of a tool which can create 'fake' HDR data, please share.
Same with upscaling, if you upscale and get new details, those are 'fake'.


Cu Selur

[davexnet]

Sure, it's 'fake' generated stuff by some machine learning stuff.
So if you know of a tool which can create 'fake' HDR data, please share.
Same with upscaling, if you upscale and get new details, those are 'fake'.


Cu Selur

My LG TV has a fake HDR picture setting. I can watch everything like this if I want; no need to do any special encoding

[pandy]

Seem Intel also provide some software solution to perform fancy pancy conversion:

Added new ffmpeg-raisr filter, which can be used by content creators and broadcasters to transform older content or user generated content created in non-digital and pre-high-definition (HD) formats into today's HD and 4K formats for use with NAS, Streaming and Video-on-Demand services

[Selur]

No me, the problem is more in the creation of the HDR data.

[pandy]

No me, the problem is more in the creation of the HDR data.

If you can colorize B/W movie then why not artificially extend light dynamic? IMHO it will be less difficult than colorization .

[Selur]

https://forum.videohelp.com/threads/414137-How-can-I-upscale-from-1080p-SDR-to-4K-HDR#post2731305 I already wrote it's possible to create 'fake' HDR data with RTX cards, but there is no tool to record those values. And I have no clue how good it really is.

[Cornucopia]

No me, the problem is more in the creation of the HDR data.

If you can colorize B/W movie then why not artificially extend light dynamic? IMHO it will be less difficult than colorization .

I disagree. Colorization has lots of existing reference color images to draw from, and certain choices are intuitive.
Where are the HDR references, and what is intuited there? We are used to understanding differences of local contrast but global contrast is new territory. Lots of room for error.


Scott

[Selur]

Lots of room for error.

Same as with automated colorization.

[pandy]

I disagree. Colorization has lots of existing reference color images to draw from, and certain choices are intuitive.
Where are the HDR references, and what is intuited there? We are used to understanding differences of local contrast but global contrast is new territory. Lots of room for error.

First - humans happily accept errors if errors suit their expectation or provide other positive feelings - average consumer like artificial, over saturated colors - sky bluer than blue, grass greener than green etc.

Secondly light is more predictable in terms of physical properties - in colorization process imaginative color can be wrong as only general knowledge is applied (like skin color can't be blue with exception of people poisoned by silver but Argyria is quite rare so person skin where luminance level is low will be rather some shade of brown/chocolate than blue/gray-blue after colorization) and calculated luminance - similar rules can be applied to light maps but light is more predictable than two colors equally probable and providing similar luminance (luminance of the dark red can be comparable to blue - and from educated guess perspective shirt can be equally dark red or light blue). We know lot of things about light as it can be described mathematically and it behave in more consistent fashion - this is of course my impression - not saying that SDR to fake-HDR conversion is easy or trivial but generally it should be not more complicated than colorization - if result looks convincing, looks appealing to customer then who cares about objective truth? (btw i hate colorization as frequently see colorized human faces or artificially pink or candle yellow - both skin colors are unnatural to me - i never passing on street people with this kind of skin color) - our civilization enjoy fake tooth's, tits, buts etc so why not fake HDR?

[sophisticles]

In addition to what has been said, allow me to share some actual examples to better illustrate the limitations of what is being asked.

I downloaded the files 2160p50 and 1080p50 created by SVT from here:

https://media.xiph.org/video/derf/

I combined the 5 clips of each resolution into a ffv1 lossless mkv and I did the following:

I encoded the 2160p50 and 1080p50 combined lossless versions to x265 crf 28 as a comparison point.

I upscaled the 1080p50 version to 2160p50 using avidemuxLinux_GLIBC_2.28_amd64_240316_1627 appimage on Manjaro.

This version contains an AI Enhance software filter for upscaling, I used the fast-FSRCNN x2 method and I also created an upscaled version using an HDR LUT.

The 1080p50 --> 2160p50 HDR file took about 25 minutes for a 50 second clip.

I also included a 1080p50 HDR clip for comparison, which took 6 minutes for the same 50 seconds.

The upscale without LUT took about 23 minutes.

The test system is a HP laptop, i5-1035G1 with 16gb ram and fast NVMe and fast SSD.

Editor's Cut of Spider-Man 3 has a run time of 2 hours 17 minutes or 8220 seconds, meaning it would take this system about 68.5 hours to upscale the whole movie using this software and this algorithm.

Math - 8220 seconds / 50 seconds = 164.4 segments 50 seconds long, if it takes 25 minutes to encode one 50 second segment then 164.4 segments should take 164.4 x 25 minutes or 4110 minutes / 60 minutes in an hour = 68 hours 30 minutes.

Is it really worth it?

[Cornucopia]

Not sure why this is designated "HDR" - doesn't look like it, is not BT.2020, doesn't seem to have any of the metadata, is 8bit. neither source nor final.


Scott

[sophisticles]

Not sure why this is designated "HDR" - doesn't look like it, is not BT.2020, doesn't seem to have any of the metadata, is 8bit. neither source nor final.

Scott

That was kind of the point, the general consensus in this thread was that it is not possible to create a true HDR from what he has.

This is supposed to be a conformation of this viewpoint.

The LUT I used is named "Presetpro - HDR Color.cube" and for some videos it does create a pleasing look.

The reality is you are not going to take Editor's Cut of Spider-Man 3 and make it look lilke the Theatrical Cut, I don't care what software you use or what hardware you use.

The best you can do is make it look a bit more appealing, depending on taste.

For instance, I assume you have played around with the 100gb version of Meridian, and possibly seen some of the re-encodes that are out there. To me, even though, or perhaps because, it was designed to be a torture test for encoders, it doesn't look visually appealing.

To that end I have been trying different ways to make it more appealing and because it's such a beast to decode and takes forever for a test encode, I created a 720p MJPEG AVI as a guinea pig and using this as source, I tried the same LUT produced the attached file, which I think looks better than the original.

People have an instinctive belief that bigger is better, more pixels must be better, and I have been guilty of this as well, but all one has to do is look at sample footage from BM's Ursa 4.6 compared to BM's original 6K camera to see that resolution is not the start and end of quality.

It would be interesting to see what could be done with something like Topaz's video and these samples.