Welcome to Doom9's Forum, THE in-place to be for everyone interested in DVD conversion.

Before you start posting please read the forum rules. By posting to this forum you agree to abide by the rules.

 

Go Back   Doom9's Forum > Hardware & Software > Software players

Reply
 
Thread Tools Search this Thread Display Modes
Old 9th October 2016, 01:28   #39641  |  Link
Georgel
Visual Novel Dev.
 
Georgel's Avatar
 
Join Date: Nov 2015
Location: Bucharest
Posts: 200
Quote:
Originally Posted by Betroz View Post
I use NNEDI3 256 for Chroma Upscaling and NNEDI3 64 for Image Doubling, and that is nice too if you got a fast GPU that can handle it (I have a 1080 card).
Wish a better algorithm than Jinc would come to image upscaling too (it's great that the algorithm for image doubling is to good), but with image upscaling, the effects are really visible on some situations.

A bit sad here that nvidia DSR doesn't work with madVR, since with most games it creates really nice results... Especially for using image upscaling of a larger factor instead of using only chroma and luma upscaling.
Georgel is offline   Reply With Quote
Old 9th October 2016, 04:46   #39642  |  Link
CarlosCaco
Registered User
 
Join Date: Oct 2015
Location: Brasil, SP, São Paulo
Posts: 154
Quote:
Originally Posted by Georgel View Post
Wish a better algorithm than Jinc would come to image upscaling too (it's great that the algorithm for image doubling is to good), but with image upscaling, the effects are really visible on some situations.

A bit sad here that nvidia DSR doesn't work with madVR, since with most games it creates really nice results... Especially for using image upscaling of a larger factor instead of using only chroma and luma upscaling.
Image doubling in madvr have superior quality than downsanpling with DSR, try supersampling.... also if you want to use DSR with madvr potplayer have an internal setting to do that (but is really a waste of resources since madvr algorithm much higher quality than DSR)
__________________
Desktop, i5 2500, 8GB, N570 GTX TF III PE/OC
Asus X555LF, i7-5500U, 6GB Ram, Nvidia 930m/HD 5500
Windows 8.1 Pro x64
CarlosCaco is offline   Reply With Quote
Old 9th October 2016, 07:38   #39643  |  Link
ryrynz
Registered User
 
ryrynz's Avatar
 
Join Date: Mar 2009
Posts: 3,646
Quote:
Originally Posted by Betroz View Post
I use NNEDI3 256 for Chroma Upscaling and NNEDI3 64 for Image Doubling, and that is nice too if you got a fast GPU that can handle it (I have a 1080 card).
Switch to 64 neurons for chroma screenshot that and 256 and compare, tell me truthfully that you think it's worth the rendering time.
PS: It's not.
ryrynz is offline   Reply With Quote
Old 9th October 2016, 07:43   #39644  |  Link
ryrynz
Registered User
 
ryrynz's Avatar
 
Join Date: Mar 2009
Posts: 3,646
Quote:
Originally Posted by Shiandow View Post
Now as far as I can tell it should look something like this instead, which doesn't seem to be possible in MadVR (yet), but maybe I'm missing something.
Maybe it's not meant to be as realistic as you'd like it to be. For comparison it may be an idea to just fire it up on a good HDR TV and compare.
ryrynz is offline   Reply With Quote
Old 9th October 2016, 08:21   #39645  |  Link
Betroz
Is this for real?
 
Betroz's Avatar
 
Join Date: Mar 2016
Location: Norway
Posts: 168
Quote:
Originally Posted by ryrynz View Post
Switch to 64 neurons for chroma screenshot that and 256 and compare, tell me truthfully that you think it's worth the rendering time.
PS: It's not.
I have tried super-xbr for Chroma Upscaling in combination with NNEDI3 128 for Image Doubling too (Luma & Chroma Doubling), but overall rendertime was a lot higher. Image Doubling is taxing as you know. So I went down to 64 neurons. Strangely enough, then I could use NNEDI3 256 for Chroma Upscaling.
__________________
My HTPC : i9 10900K | nVidia RTX 4070 Super | TV : Samsung 75Q9FN QLED
Betroz is offline   Reply With Quote
Old 9th October 2016, 09:17   #39646  |  Link
huhn
Registered User
 
Join Date: Oct 2012
Posts: 7,903
Quote:
Originally Posted by CarlosCaco View Post
Image doubling in madvr have superior quality than downsanpling with DSR, try supersampling.... also if you want to use DSR with madvr potplayer have an internal setting to do that (but is really a waste of resources since madvr algorithm much higher quality than DSR)
the biggest problem is that people don't understand the difference between DSR using on a game and DSR using on a video.

a video has a fixed resolution most PC games doesn't have a fixed resolution.

taking this into account it should be very very easy to understand why DSR has a potential positive effect on picture quality and not on video picture quality.
huhn is offline   Reply With Quote
Old 9th October 2016, 10:50   #39647  |  Link
Asmodian
Registered User
 
Join Date: Feb 2002
Location: San Jose, California
Posts: 4,406
Quote:
Originally Posted by Betroz View Post
I have tried super-xbr for Chroma Upscaling in combination with NNEDI3 128 for Image Doubling too (Luma & Chroma Doubling), but overall rendertime was a lot higher. Image Doubling is taxing as you know. So I went down to 64 neurons. Strangely enough, then I could use NNEDI3 256 for Chroma Upscaling.
Even if you have the performance available 256 neuron chroma upscaling is a waste of electricity/heat. I suggest you never use above 64 neurons for chroma upscaling, even that is overkill but less overkill than 256 neurons, which is just excessive.

Luma/Chroma doubling is harder than chroma upscaling because, assuming 1080p 4:2:0 video, chroma upscaling has to scale 960x540 to 1920x1080 while luma/chroma doubling has to scale 1920x1080 to 3840x2160. Also chroma doubling is twice as hard as luma doubling; chroma is two planes, U and V, while luma is only one, Y. Your results are not strange at all.

Use as high as you can for luma doubling but don't go above 32 neurons for chroma doubling and 64 neurons for chroma upscaling.

With a Titan X (Pascal) I have plenty of performance available but using more neurons for chroma really is pointless.
__________________
madVR options explained
Asmodian is offline   Reply With Quote
Old 9th October 2016, 11:26   #39648  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by Shiandow View Post
Has anyone who fiddled around with mapping HDR to an SDR screen figured out a good setting for 'hue preservation'? I've been doing some tests with Sony's Camp sample, and I can't make figure out a way to make the fire at 1:09 look vaguely realistic. The best I can manage is just to turn hue preservation off, which makes it look like this.

Now as far as I can tell it should look something like this instead, which doesn't seem to be possible in MadVR (yet), but maybe I'm missing something.
We've had discussions about the best way to handle these situations in this thread a while ago. I've also talked to Calman, Florian (displaycal), zoyd and Graeme about this.

The key is the "fix too bright & saturated pixels by" combobox in the madVR settings. Try setting that to "0% luminance reduction", and you'll get something nearer to what you expect. This may produce better looking results for this specific scene, but in other scenes it looks worse because the pixels actually become near white which means they're losing their color information.

You can create HDR conversion 3dluts with the latest dispcal. The 3dluts seem to be slightly better than what madVR does with pixel shaders, but the overall look is pretty similar.
madshi is offline   Reply With Quote
Old 9th October 2016, 11:37   #39649  |  Link
AngelGraves13
Registered User
 
Join Date: Dec 2010
Posts: 254
Quote:
Originally Posted by Betroz View Post
I use NNEDI3 256 for Chroma Upscaling and NNEDI3 64 for Image Doubling, and that is nice too if you got a fast GPU that can handle it (I have a 1080 card).
I have a 1080 GTX as well (FTW) and Jinc with SR 4 for Chroma and Scaling with Linear Light and Anti-ringing checked looks much better than NNEDI3 256. The image looks real, not digitized. SuperXBR and NNEDI3 have a very digitized and processed look to them.

NNEDI3 is a good scaler for chroma, but it makes colors look overprocessed and in some cases they can cross.

Jinc is the cleanest scaler around and when you super-res it it only gets sharper. Other scalers like NNEDI3 and SuperXBR gets sharper too, but you start to see the artifacts easily.
AngelGraves13 is offline   Reply With Quote
Old 9th October 2016, 12:48   #39650  |  Link
Sunset1982
Registered User
 
Join Date: Sep 2014
Posts: 280
Quote:
Originally Posted by AngelGraves13 View Post
I have a 1080 GTX as well (FTW) and Jinc with SR 4 for Chroma and Scaling with Linear Light and Anti-ringing checked looks much better than NNEDI3 256. The image looks real, not digitized. SuperXBR and NNEDI3 have a very digitized and processed look to them.

NNEDI3 is a good scaler for chroma, but it makes colors look overprocessed and in some cases they can cross.

Jinc is the cleanest scaler around and when you super-res it it only gets sharper. Other scalers like NNEDI3 and SuperXBR gets sharper too, but you start to see the artifacts easily.

So what would you suggest for upscaling to 4K?

Chroma: NNEDI3
Image: Jinc +LL +AR +SR4

Atm I use:

Chroma: SuperXBR 100, AR, SR1
Image Doubling: SuperXBR AB25
Image: Jinc AR
__________________
Intel i5 6600, 16 GB DDR4, AMD Vega RX56 8 GB, Windows 10 x64, Kodi DS Player 17.6, MadVR (x64), LAV Filters (x64), XySubfilter .746 (x64)
LG 4K OLED (65C8D), Denon X-4200 AVR, Dali Zensor 5.1 Set
Sunset1982 is offline   Reply With Quote
Old 9th October 2016, 13:11   #39651  |  Link
Betroz
Is this for real?
 
Betroz's Avatar
 
Join Date: Mar 2016
Location: Norway
Posts: 168
Quote:
Originally Posted by Asmodian View Post
Even if you have the performance available 256 neuron chroma upscaling is a waste of electricity/heat. I suggest you never use above 64 neurons for chroma upscaling, even that is overkill but less overkill than 256 neurons, which is just excessive.

Luma/Chroma doubling is harder than chroma upscaling because, assuming 1080p 4:2:0 video, chroma upscaling has to scale 960x540 to 1920x1080 while luma/chroma doubling has to scale 1920x1080 to 3840x2160. Also chroma doubling is twice as hard as luma doubling; chroma is two planes, U and V, while luma is only one, Y. Your results are not strange at all.

Use as high as you can for luma doubling but don't go above 32 neurons for chroma doubling and 64 neurons for chroma upscaling.

With a Titan X (Pascal) I have plenty of performance available but using more neurons for chroma really is pointless.
Thanks for your advice, I'll check it out
__________________
My HTPC : i9 10900K | nVidia RTX 4070 Super | TV : Samsung 75Q9FN QLED
Betroz is offline   Reply With Quote
Old 9th October 2016, 13:11   #39652  |  Link
Ver Greeneyes
Registered User
 
Join Date: May 2012
Posts: 447
Quote:
Originally Posted by Betroz View Post
Strangely enough, then I could use NNEDI3 256 for Chroma Upscaling.
It's not that strange: the chroma resolution is usually a quarter of the luma resolution (half the height and half the width), but there are two chroma channels and only one luma channel - so NNEDI3 256 for chroma upscaling is comparable to NNEDI3 128 for luma doubling (as doubling the number of neurons roughly doubles the cost).

So while NNEDI3 256 for chroma upscaling still costs more than NNEDI3 64 for luma, going down to NNEDI3 64 for luma must have freed up enough resources to use NNEDI3 256 for chroma upscaling.

Last edited by Ver Greeneyes; 9th October 2016 at 13:16.
Ver Greeneyes is offline   Reply With Quote
Old 9th October 2016, 13:14   #39653  |  Link
Betroz
Is this for real?
 
Betroz's Avatar
 
Join Date: Mar 2016
Location: Norway
Posts: 168
Quote:
Originally Posted by AngelGraves13 View Post
I have a 1080 GTX as well (FTW) and Jinc with SR 4 for Chroma and Scaling with Linear Light and Anti-ringing checked looks much better than NNEDI3 256. The image looks real, not digitized. SuperXBR and NNEDI3 have a very digitized and processed look to them.

NNEDI3 is a good scaler for chroma, but it makes colors look overprocessed and in some cases they can cross.

Jinc is the cleanest scaler around and when you super-res it it only gets sharper. Other scalers like NNEDI3 and SuperXBR gets sharper too, but you start to see the artifacts easily.
From my testing Jinc AR + super-rez is not as sharp as NNEDI3 64 and higher, but I didn't try SR4, only SR2 in my testing.
__________________
My HTPC : i9 10900K | nVidia RTX 4070 Super | TV : Samsung 75Q9FN QLED
Betroz is offline   Reply With Quote
Old 9th October 2016, 13:16   #39654  |  Link
Betroz
Is this for real?
 
Betroz's Avatar
 
Join Date: Mar 2016
Location: Norway
Posts: 168
Quote:
Originally Posted by Ver Greeneyes View Post
It's not that strange: the chroma resolution is usually a quarter of the luma resolution (half the height and half the width), so NNEDI3 256 for chroma upscaling is comparable to NNEDI3 64 for luma doubling (as doubling the number of neurons roughly doubles the cost).

So by going from NNEDI3 128 to NNEDI3 64 for luma, you freed up roughly enough resources to do NNEDI3 256 for chroma.
Yeah that makes sense. I use NNEDI3 64 for Chroma doubling too. But will try out the settings Asmodian adviced me to use.
__________________
My HTPC : i9 10900K | nVidia RTX 4070 Super | TV : Samsung 75Q9FN QLED
Betroz is offline   Reply With Quote
Old 9th October 2016, 13:18   #39655  |  Link
Ver Greeneyes
Registered User
 
Join Date: May 2012
Posts: 447
Quote:
Originally Posted by Betroz View Post
Yeah that makes sense. I use NNEDI3 64 for Chroma doubling too. But will try out the settings Asmodian adviced me to use.
I forgot that chroma has 2 channels, updated my post above. Either way, differences in quality from chroma upscaling are often so negligible that it's worth more to put those resources into luma doubling
Ver Greeneyes is offline   Reply With Quote
Old 9th October 2016, 14:00   #39656  |  Link
Shiandow
Registered User
 
Join Date: Dec 2013
Posts: 753
Quote:
Originally Posted by madshi View Post
We've had discussions about the best way to handle these situations in this thread a while ago. I've also talked to Calman, Florian (displaycal), zoyd and Graeme about this.

The key is the "fix too bright & saturated pixels by" combobox in the madVR settings. Try setting that to "0% luminance reduction", and you'll get something nearer to what you expect. This may produce better looking results for this specific scene, but in other scenes it looks worse because the pixels actually become near white which means they're losing their color information.
I think I may have figured out why I didn't like that result much. It preserves the hue, which gives it a rather unnatural look. At least I've never seen a fire look anything like this before.

If instead you tried to pick the closest RGB colour, while preserving the luminance you'd get something like this, which already looks a bit more like an actual fire. Now if you were to pick the colour closest in Lab I think you'd get something like this, but I'm not still not entirely sure if the method I used is 100% correct.

You're right that you'll lose some colour information this way, but you're going to lose some information no matter what you do, and I'm not convinced that hue and saturation are more important than luminance.
Shiandow is offline   Reply With Quote
Old 9th October 2016, 14:13   #39657  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by Shiandow View Post
I think I may have figured out why I didn't like that result much. It preserves the hue, which gives it a rather unnatural look. At least I've never seen a fire look anything like this before.

If instead you tried to pick the closest RGB colour, while preserving the luminance you'd get something like this, which already looks a bit more like an actual fire. Now if you were to pick the colour closest in Lab I think you'd get something like this, but I'm not still not entirely sure if the method I used is 100% correct.
FWIW, Lab does not really preserve hue angles well. Lab is better than gamma or linear light RGB, but still far from optimal for this kind of processing.

Quote:
Originally Posted by Shiandow View Post
You're right that you'll lose some colour information this way, but you're going to lose some information no matter what you do, and I'm not convinced that hue and saturation are more important than luminance.
I've started out with the same thinking, and put luminance preservation at highest priority. But it results in serious problems in many situations. E.g. on a quick search I found this:

http://forum.doom9.org/showpost.php?...ostcount=38653

If you put priority on reproducing luminance correctly, the orange headlight of the truck becomes white. That looks really bad in that specific image.

Have you played with the "fix too bright & saturated pixels by" setting in madVR? It allows you to decide for yourself if you value luminance higher than hue/saturation.
madshi is offline   Reply With Quote
Old 9th October 2016, 15:05   #39658  |  Link
Shiandow
Registered User
 
Join Date: Dec 2013
Posts: 753
Quote:
Originally Posted by madshi View Post
FWIW, Lab does not really preserve hue angles well. Lab is better than gamma or linear light RGB, but still far from optimal for this kind of processing.
Well, if I did everything correctly I should be able to minimize the distance in any colour space. Unfortunately, for most spaces more advanced than Lab the difference in colour isn't measured as a simple distance, which complicates things slightly. Something like CIE94 might just be possible, although it's not symmetric, which is odd. Worth a try anyway.

Quote:
Originally Posted by madshi View Post
I've started out with the same thinking, and put luminance preservation at highest priority. But it results in serious problems in many situations. E.g. on a quick search I found this:

http://forum.doom9.org/showpost.php?...ostcount=38653

If you put priority on reproducing luminance correctly, the orange headlight of the truck becomes white. That looks really bad in that specific image.
That headlight does seem to be over 2000 nits bright though, on a 400~600 nit screen 'white' is probably the closest you're going to get, in any colour space.

The real problem seems to be that simply picking the closest RGB colour will lead to some lost detail, regardless of how you define distances. Fixing that probably can't be done by looking at individual pixels.

Quote:
Originally Posted by madshi View Post
Have you played with the "fix too bright & saturated pixels by" setting in madVR? It allows you to decide for yourself if you value luminance higher than hue/saturation.
Well, seems I value luminance higher than saturation, and saturation higher than hue. Which is unfortunate since the one thing your method does perfectly is preserving the hue, at the cost of luminance and saturation.
Shiandow is offline   Reply With Quote
Old 9th October 2016, 15:21   #39659  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by Shiandow View Post
Well, seems I value luminance higher than saturation, and saturation higher than hue. Which is unfortunate since the one thing your method does perfectly is preserving the hue, at the cost of luminance and saturation.
Have you tried setting the "fix too bright & saturated pixels by" option to "0% luminance reduction"? That should do what you want?
madshi is offline   Reply With Quote
Old 9th October 2016, 15:48   #39660  |  Link
Shiandow
Registered User
 
Join Date: Dec 2013
Posts: 753
Quote:
Originally Posted by madshi View Post
Have you tried setting the "fix too bright & saturated pixels by" option to "0% luminance reduction"? That should do what you want?
That resulted in this image, which I think looks rather unnatural. Now maybe it's an accident that fire looks better if you don't preserve hue, but I don't really know how to test that (well an HDR screen might help, but if I had one I wouldn't be trying to figure out how to display HDR on an SDR screen).
Shiandow is offline   Reply With Quote
Reply

Tags
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling

Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off

Forum Jump


All times are GMT +1. The time now is 10:01.


Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2024, vBulletin Solutions Inc.