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 20th July 2015, 23:00   #31981  |  Link
JarrettH
Registered User
 
Join Date: Aug 2004
Location: Canada
Posts: 860
I love the smooth contours of super-xbr

I think we've gone a little overboard obsessed with sharpening. Doubling already increases sharpness and super-xbr needs no enhancement IMO.

Would it be possible to add super-xbr at 75 to this comparison?

Super-xbr definitely looks sharp, and when you compare it to the original, it retains the blinding white brightness reflected from the castle.
JarrettH is offline   Reply With Quote
Old 20th July 2015, 23:09   #31982  |  Link
Gagorian
Registered User
 
Join Date: Jul 2013
Posts: 27
madshi are you familiar with the DarbeeVision Darblet used amongst HT enthusiasts? I've been quite happy with the sharpening it provides (at low strength). I wonder if there's any similar algorithm available to run on the GPU?
Gagorian is offline   Reply With Quote
Old 20th July 2015, 23:24   #31983  |  Link
har3inger
Registered User
 
Join Date: Feb 2014
Posts: 139
Thanks very much for providing that comparison post, madshi. It gives us a better idea of what to look for from SuperRes and must have taken a lot of time to write up.
har3inger is offline   Reply With Quote
Old 20th July 2015, 23:41   #31984  |  Link
baii
Registered User
 
Join Date: Dec 2011
Posts: 180
I seem to get periodic frame drop (by either 1 or 2 frame on osd) when playing interlaced material in dx11 windowed mode. I don't seem to notice all(or any) the frame drop by eye(which is normal I think, consider it is 60fps video.., don't have that kind of eye xd). Would this be a osd glitch or real frame drop?
It is fine in dx11 fse or dx9.
I had play with all the vync/seperate device options, but none helps.
Configuration is win8.1, Amd dxva de-interlance, reclock, 30i material, 59/60hz screen. Tested in recent .14 and .20.

Sent from my 306SH

Last edited by baii; 20th July 2015 at 23:44.
baii is offline   Reply With Quote
Old 20th July 2015, 23:51   #31985  |  Link
ryrynz
Registered User
 
ryrynz's Avatar
 
Join Date: Mar 2009
Posts: 3,646
Quote:
Originally Posted by JarrettH View Post
I think we've gone a little overboard obsessed with sharpening.
Indeed. I'm waiting for denoising TBH.
ryrynz is offline   Reply With Quote
Old 21st July 2015, 00:00   #31986  |  Link
har3inger
Registered User
 
Join Date: Feb 2014
Posts: 139
I have found a bug with 30 fps content with the latest .20 build. OSD reported render times are well under 2 ms (which is impossible for my hardware specs: last version reports render times ~28ms for same settings), but cranking up settings will caused dropped frames while the OSD reports <20ms render times. This is a cosmetic bug with no functional problems that I can see. Problem didn't occur in .19 and earlier. No rendering time calculation errors with 24 fps content. Don't have access to any other framerate content to test.

Windows 8.1 x64, mpc-hc x64, madvr x64, LAV filters 65.0. Image upscale jinc ar, chroma bicubic 75 ar. No image doubling.
har3inger is offline   Reply With Quote
Old 21st July 2015, 00:05   #31987  |  Link
Della
Registered User
 
Join Date: Dec 2011
Posts: 32
Outstanding post Madshi. I now have a clear, visual understanding of the various options.
Thank you.
Della is offline   Reply With Quote
Old 21st July 2015, 00:51   #31988  |  Link
pirlouy
_
 
Join Date: May 2008
Location: France
Posts: 692
You may have lost a lot of time writing this, but for a lurker, this is the kind of post I like to read. Images for us to judge, and your comments/POV on this. Definitively a good post. It avoids endless discussions without proof. And it allows us to choose accordingly.
pirlouy is offline   Reply With Quote
Old 21st July 2015, 01:21   #31989  |  Link
MS-DOS
Registered User
 
Join Date: Sep 2012
Posts: 77
Thanks for the great post, madshi.

However, using 256 neurons for NNEDI3 as "the best option" may not be a good idea. I've recently found an example where 256 neurons cause a very notable artifact:

Original -|- NNEDI3 128 -|- NNEDI3 256

See a tiny white vertical line above Ramzesss word right between "m" and "z" ? It's not present on the original image, nor with any neuron count below 256.
MS-DOS is offline   Reply With Quote
Old 21st July 2015, 01:40   #31990  |  Link
MistahBonzai
Registered User
 
Join Date: Mar 2013
Posts: 101
Quote:
Originally Posted by madshi View Post
I've decided to make a big screenshot comparison post to compare the different upscaling and sharpening / post-processing options we have now. In order to make all this more "scientific", I've decided that we need a "ground truth" image to compare to, because that's the only way to properly judge if an algorithm produces good results or not...

-------

Comments welcome!
That was Crème de la Chem to this OCD pixel peeping vidiot! I really appreciate your dedication and the hard work you put into this project. And that was another application of icing on the cake. Yup, truly outstanding.
MistahBonzai is offline   Reply With Quote
Old 21st July 2015, 02:11   #31991  |  Link
Hyllian
Registered User
 
Hyllian's Avatar
 
Join Date: Jun 2014
Posts: 42
Quote:
Originally Posted by madshi View Post
I've decided to make a big screenshot comparison post to compare the different upscaling and sharpening / post-processing options we have now. In order to make all this more "scientific", I've decided that we need a "ground truth" image to compare to, because that's the only way to properly judge if an algorithm produces good results or not. If we don't have a ground truth to compare to, all our evaluations are somewhat subjective. So here comes:

Bilinear+SuperRes -|- Jinc+SuperRes -|- super-xbr+SuperRes -|- NNEDI3+SuperRes -|- GroundTruth


-------

Comments welcome!
Congrats for the amazing post! I confess I was a non-believer of SuperRes, but now I can see the light! It's almost as good as waifu2x-castle.

It's just me or the sxbr+sres = nnedi3+sres? I can't spot a difference in a still, so in motion it should be even harder.

How many passes did you use for SuperRes in these examples?

EDIT: even bilinear+sres is almost identical to the others, with slightly worse aliasing!

Last edited by Hyllian; 21st July 2015 at 02:42.
Hyllian is offline   Reply With Quote
Old 21st July 2015, 02:54   #31992  |  Link
RyuzakiL
Registered User
 
Join Date: May 2015
Posts: 36
Quote:
Originally Posted by Hyllian View Post
Congrats for the amazing post! I confess I was a non-believer of SuperRes, but now I can see the light! It's almost as good as waifu2x-castle.

It's just me or the sxbr+sres = nnedi3+sres? I can't spot a difference in a still, so in motion it should be even harder.

How many passes did you use for SuperRes in these examples?

EDIT: even bilinear+sres is almost identical to the others, with slightly worse aliasing!

see? that's what i've been using for (S-XBR+SuperRes) ever since the feature got introduced. I don't know why there are these NEDDI3 256 Fanatics/cult out there who always claim the holy grail for Picture Quality when the Madshi's Screenshots speaks for themselves. When performance is still important before Picture Quality (where NNEDI3 really lacks and burns precious GPU resources).

S-XBR+SuperRes = NNEDI3+SuperRes but very light on GPU load, Ergo you can build a HTPC with a cheaper GPU (less heat and power consumption) but at the same time still can use better MAdvr settings like this time (S-XBR+SuperRes).

Last edited by RyuzakiL; 21st July 2015 at 02:58.
RyuzakiL is offline   Reply With Quote
Old 21st July 2015, 03:30   #31993  |  Link
leeperry
Kid for Today
 
Join Date: Aug 2004
Posts: 3,477
TYVM for the comparisons, I'll try my luck this week with running the same procedure but with NNEDI16/32 and 3@0.42 LQ SR in .15 coz comparing NNEDI256 to über-sharp LQ SR in .19 is kinda putting things out of context IMVHO, only one way to find out anyway and your testing procedure makes a heck lot of sense indeed gg

You don't seem too keen on telling us what SR does in mVR but it's not a "dumb" sharpener as it would appear? Does it take previous frames in account like the vmotion stuff and as the name implies?

Last edited by leeperry; 21st July 2015 at 03:47.
leeperry is offline   Reply With Quote
Old 21st July 2015, 03:31   #31994  |  Link
RyuzakiL
Registered User
 
Join Date: May 2015
Posts: 36
Bilateral+JincAR+SuperRes4/0.66 from 1080pHQ Source

It seems the said config is very good PQ/Performance wise for processing 1080pHQ Source.

I'm really interested on doing Madshi's Format of Picture/Screenshot comparison series.

But for the life of me, i cannot grab a screenshot on DX11 FSE10bit mode?

or i should switch to DX11 Windowed FS 8bit mode?
RyuzakiL is offline   Reply With Quote
Old 21st July 2015, 03:45   #31995  |  Link
leeperry
Kid for Today
 
Join Date: Aug 2004
Posts: 3,477
Quote:
Originally Posted by RyuzakiL View Post
But for the life of me, i cannot grab a screenshot on DX11 FSE10bit mode?
Yeah I'm also not too sure how to make screenshots in PotP, you'll apparently need to disable exclusive mode and press the <printscreen> key on your keyboard.
leeperry is offline   Reply With Quote
Old 21st July 2015, 05:10   #31996  |  Link
MistahBonzai
Registered User
 
Join Date: Mar 2013
Posts: 101
Quote:
Originally Posted by leeperry View Post
Yeah I'm also not too sure how to make screenshots in PotP, you'll apparently need to disable exclusive mode and press the <printscreen> key on your keyboard.
Yup - black images are not that informative

If working with MPC-HC these procedures will enable you to capture specific frames. I have to believe that there is a way to seek by frame in most video players..

This procedure saves the onscreen image assuming you are *not* in exclusive mode.
  1. Uncheck "enable automatic fullscreen exclusive mode" in MadVR "general settings".
  2. Open video in MPC-HC.
  3. Pause video.
  4. <Ctrl>+<g>
  5. specify desired frame - example Frame: (2000, 23.976)
  6. <alt>+<print screen>
  7. Paste into video editor.

This procedure saves the image from the MPC-HC image renderer and not from MadVR
  1. Open video in MPC-HC.
  2. Pause video.
  3. <Ctrl>+<G>
  4. specify desired frame - example Frame: (2000, 23.976)
  5. <Alt>+<i>
  6. Select either jpg or png.
  7. Save to the perefered folder.
Note: More than anything I would like a way to do MadVR screen shots from MPC-HC

Last edited by MistahBonzai; 21st July 2015 at 05:28.
MistahBonzai is offline   Reply With Quote
Old 21st July 2015, 08:05   #31997  |  Link
James Freeman
Registered User
 
Join Date: Sep 2013
Posts: 919
Quote:
Originally Posted by madshi View Post
Quote:
D3D11,FSE unreported stutter
FWIW, can you try setting the flush setting for "after intermediate render steps" to "flush & wait (sleep)"? Does that help? Probably not, but it would be important to know for me.
Nope. Still stutters.
__________________
System: i7 3770K, GTX660, Win7 64bit, Panasonic ST60, Dell U2410.
James Freeman is offline   Reply With Quote
Old 21st July 2015, 08:11   #31998  |  Link
QBhd
QB the Slayer
 
QBhd's Avatar
 
Join Date: Feb 2011
Location: Toronto
Posts: 697
For PotPlayer Video==>Video Capture.... Or just use Ctrl+Alt+E to grab the screen frame. Might want to go the menu route first to make sure you save as bmp or png. Ctrl+E will grab the source frame.

QB
__________________
QBhd is offline   Reply With Quote
Old 21st July 2015, 09:02   #31999  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Thanks for the positive feedback, guys! Happy to hear you like my comparison post!

Quote:
Originally Posted by Akeno View Post
Adaptive Sharpen: Seems pretty destructive when it comes to real life material. It works a lot better for anime material where there aren't as many fine details and is the most ideal image enhancer.
Yeah, I wasn't meaning to say that AdaptiveSharpen should never be used. It could be useful to sharpen overly soft sources, or simply to add some "pop". However, "Upscaling Refinement" was always planned by me to post-process an upscaled image in such a way that it gets nearer to the ground truth. So, I would say that for this specific purpose I would not recommend AdaptiveSharpen.

I'm not fully sure yet if the split with "image enhancements" and "upscaling refinement" will stay as it is. I've had the idea that I could probably modify FineSharp and AdaptiveSharpen so that they have the same effectiveness/strength when applied after upscaling compared to before upscaling. Sharpening after upscaling generally looks better. So it's possible the "image enhancements" may run after scaling in a future version, too (maybe with a trade quality option to perform it before scaling). But I don't know for sure yet.

Quote:
Originally Posted by Akeno View Post
FineSharp: Feels unusable at this point due to the extreme artifacts it produces, although I suppose there would be no problem with a nearly 100% clean source.
IIRC Didée also wrote it with very clean sources in mind. IMHO FineSharp does an amazing job considering its speed and simplicity. But requiring a perfectly clean source is of course problematic.

Quote:
Originally Posted by Akeno View Post
LumaSharpen: I can't see a reason to use it personally. It just isn't as noticeable as the other algorithms but maybe my settings are wrong.
There's a reason why I didn't include it in the comparison. LumaSharpen is very fast, but other than that I wouldn't recommend to use it.

Quote:
Originally Posted by Akeno View Post
On a side note, madshi, you recommend 32 Neurons as a minimum for luma doubling due to artifacts with 16. My system can't run 32 with 30fps material but I haven't noticed anything wrong with 16. Could you elaborate on what kind of artifacts there are or post some pictures?
Well, 16 neurons is ok, too, for many sources. It's just that it has a lot more aliasing in some situations (edge angles?) compared to 32. It depends on the situation, though. Often 16 neurons is just as good as 32 neurons.

Quote:
Originally Posted by JarrettH View Post
I think we've gone a little overboard obsessed with sharpening. Doubling already increases sharpness and super-xbr needs no enhancement IMO.
Quote:
Originally Posted by ryrynz View Post
Indeed. I'm waiting for denoising TBH.
Well, you guys would probably like it if madVR could magically turn any downscaled image back into its original high res image, wouldn't you? So that's exactly what I'm aiming at. If you look at the screenshots, the "sharpening" applied by SuperRes brings us nearer to the ground truth, which can only be considered a good thing, no?

That said, yes, some algorithms to clean up bad sources would be great. But I prefer to do one thing right first, before switching to the next thing.

Quote:
Originally Posted by JarrettH View Post
Would it be possible to add super-xbr at 75 to this comparison?
Possible, yes, but I don't think it's worth the effort. super-xbr 75 is similar to super-xbr 100, just a bit softer with some less ringing. Should make no difference to directional artifacts.

Quote:
Originally Posted by Gagorian View Post
madshi are you familiar with the DarbeeVision Darblet used amongst HT enthusiasts? I've been quite happy with the sharpening it provides (at low strength). I wonder if there's any similar algorithm available to run on the GPU?
I've not seen Darbee myself yet. But I've been told it does this:

> It's a frame-delayed unsharp mask. So if there's no
> movement in the scene, it's just plain unsharp mask.
> If there's a lot of movement, it's a random artifact
> generator. If there's a little movement, like on a slow
> pan or a helicopter shot, it's an unsharp mask that is
> more aggressive on one side of a detail and less
> aggressive on the other side, adding a little shadow.

So for still images, it should be somewhat similar to AdaptiveSharpen. One big disadvantage of current Darbee implementations is that if you let the display upscale (e.g. 4K display), Darbee is applied before upscaling. That's like applying AdaptiveSharpen in "image enhancements". It's effective, but ugly. You'd get better quality doing it after upscaling. But AFAIK current Darbee implementations don't have the power to do 4K.

Anyway, personally I'm not really a fan of the look that unsharp mask style sharpening produces. To my eyes it looks like fat lines get even fatter. It seems to "bloat up" the image somehow, adding more contrast and pop, but making the image look more artificial. You can see some of that in my screenshot comparison post when comparing AdaptiveSharpen to FineSharp. The look which AdaptiveSharpen produces is typical for unsharp mask style sharpening.

Quote:
Originally Posted by baii View Post
I seem to get periodic frame drop (by either 1 or 2 frame on osd) when playing interlaced material in dx11 windowed mode. I don't seem to notice all(or any) the frame drop by eye(which is normal I think, consider it is 60fps video.., don't have that kind of eye xd). Would this be a osd glitch or real frame drop?
It is fine in dx11 fse or dx9.
I had play with all the vync/seperate device options, but none helps.
Configuration is win8.1, Amd dxva de-interlance, reclock, 30i material, 59/60hz screen. Tested in recent .14 and .20.
So it occurs with .14, too? Might be that your GPU is near its power limit, and using DX11 windowed mode just pushes it over the edge sometimes. Don't know for sure. It's probably a real drop. You might not see it during scenes where there's not a lot of motion. Is it a natively interlaced source (e.g. sports)? Or is it telecined film? In the latter case you could force madVR's film mode. That way GPU power reduces by a factor of 2.5x, while producing smoother playback (no 3:2 judder).

Quote:
Originally Posted by MS-DOS View Post
However, using 256 neurons for NNEDI3 as "the best option" may not be a good idea. I've recently found an example where 256 neurons cause a very notable artifact:

Original -|- NNEDI3 128 -|- NNEDI3 256

See a tiny white vertical line above Ramzesss word right between "m" and "z" ? It's not present on the original image, nor with any neuron count below 256.
Yes, I see it. I'm not sure if 256 neurons are generally a bad idea, though. NNEDI3 is not artifact free. In my experience any neuron number can produce artifacts in some images, where a different neuron count might not produce an artifact. There might be situations where 128 neurons produce an artifact and 256 neurons do not. Anyway, 256 neurons is so expensive, most users won't have the GPU power for that, anyway. In any case, it's probably better to use a much lower neuron count and run some SuperRes passes afterwards.

Quote:
Originally Posted by Hyllian View Post
I confess I was a non-believer of SuperRes, but now I can see the light! It's almost as good as waifu2x-castle.
waifu2x is quite amazing. It does clearly beat NNEDI256+SuperRes, when looking at the geometrical features in the castle image. However, I think SuperRes reproduces the trees better. All those "clever" upscaling algorithms often mess up trees and nature. SuperRes somehow manages to reproduce a more natural look there. Would be interesting to see how waifu2x+SuperRes would look like. With a bit of luck, it might combine the best of NNEDI3+SuperRes and waifu2x.

Anyway, Shiandow has estimated that waifu2x is probably at least 100x slower than NNEDI3-256, maybe more. So it's not suitable for real time video playback in the next 7 years at least, I would say.

Quote:
Originally Posted by Hyllian View Post
It's just me or the sxbr+sres = nnedi3+sres? I can't spot a difference in a still, so in motion it should be even harder.

How many passes did you use for SuperRes in these examples?
It's not just you. They're almost identical. There's one artifact, though, which super-xbr has (and NNEDI3 has not), which SuperRes doesn't fully remove. See red circle here:

http://madVR.com/doom9/temp/CastleSuperXbr.png

This specific geometrical configuration seems to be difficult. Even waifu2x has problems with this. Not on this specific tower, though. But if you look at the smaller right most tower, waifu2x is not perfect with the same geometrical configuration there. NNEDI3 handles that better, but NNEDI3 has small problems in other areas.

The screenshots were created with 8 SuperRes passes.

Quote:
Originally Posted by leeperry View Post
TYVM for the comparisons, I'll try my luck this week with running the same procedure but with NNEDI16/32 and 3@0.42 LQ SR in .15 coz comparing NNEDI256 to über-sharp LQ SR in .19 is kinda putting things out of context IMVHO, only one way to find out anyway and your testing procedure makes a heck lot of sense indeed gg
Glad to hear you like my testing procedure. I tried to find a way to make things less subjective, to increase the chances that we may end up on the same page.

Quote:
Originally Posted by leeperry View Post
You don't seem too keen on telling us what SR does in mVR but it's not a "dumb" sharpener as it would appear? Does it take previous frames in account like the vmotion stuff and as the name implies?
No, it only looks at one frame at a time. It pretty much does what Shiandow explained in that post you quoted earlier. But Shiandow's explanation was for the old SuperRes algo (where we still had number options for anti-aliasing and anti-ringing etc). The new algorithm is based on some newer scientific papers. It works like this:

0) Let e.g. super-xbr upscale the OI (original image). Result: UI (upscaled image).
1) Downscale the UI back to the same size as the OI (using: LQ=Bilinear; HQ=Bicubic).
2) Substract the OI from the downscaled UI, which gives us a sort of "error" image (EI) in the size of the OI.
3) Modify the UI in a clever way, by making use of the EI, to reduce the error.

Steps 1) to 3) is one SuperRes pass. If you want 2 SuperRes passes, you simply perform 1,2,3) twice.

All the magic happens in 3). And this code totally changed from the old to the new SuperRes algorithm. The other parts mostly remained unchanged. The old SuperRes code introduced aliasing and ringing, which is why Shiandow had added extra anti-aliasing and anti-ringing algos. The new SuperRes algorithm doesn't (or shouldn't) produce aliasing, anymore, and not as much ringing as the old code, so that's why there's no anti-ringing and anti-aliasing code in the new algorithm, anymore. I did add my own anti-ringing code in, though.

Quote:
Originally Posted by RyuzakiL View Post
I'm really interested on doing Madshi's Format of Picture/Screenshot comparison series.
Yes, please!

I've done all stuff in windowed mode. I've selected a ground truth image which in 100% view exactly fit in the media player in windowed mode, to make things easier for me. The ground truth is already a downscaled image from an even larger original image. So I could choose the exact size I wanted to have for the ground truth.

Please make sure you downscale using madVR with AR (anti-ringing). Don't downscale using an image/photo editor, unless it also has AR (I'm not aware of any that has).

Quote:
Originally Posted by James Freeman View Post
Nope. Still stutters.
Ok, thanks.

Last edited by madshi; 21st July 2015 at 09:46.
madshi is offline   Reply With Quote
Old 21st July 2015, 09:02   #32000  |  Link
Braum
Registered User
 
Join Date: Jan 2015
Posts: 48
Quote:
Originally Posted by leeperry View Post
Yeah I'm also not too sure how to make screenshots in PotP, you'll apparently need to disable exclusive mode and press the <printscreen> key on your keyboard.
It works that way, but you can also setup your own keybord shortcuts (but still need to disable FSE :/).

Braum 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 01:36.


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