madVR - high quality video renderer (GPU assisted)

[James Freeman]

Yes, I noticed this too, although not the correlation with doing stuff on the primary monitor (I didn't look too close though). I noticed that with DX11 mode, I would get bunches of dropped frames periodically (like, a bunch in a row), and sometimes the queues would fail to fill until I paused and unpaused. With DX9, the presentation glitches counter would steadily rise (causing visual glitches), but it wouldn't actually drop any frames on madVR's end.

I have no idea whether or not this is a regression though - the presentation glitches seem like a DWM problem, but perhaps the DX11 behavior could be improved.

Exactly.
Please try to recreate this behavior by scrolling in browser on the primary monitor while wathching fullscreen on the secondary.
I may have to test erlier builds to help madshi pinpoint the origin of the problem.
Anyone else use a secondary display to verify this?

[Anime Viewer]

Exactly.
Please try to recreate this behavior by scrolling in browser on the primary monitor while wathching fullscreen on the secondary.
I may have to test erlier builds to help madshi pinpoint the origin of the problem.
Anyone else use a secondary display to verify this?

I usually run with video playing on my secondary display while occasionally browsing or typing in a browser window on my primary (in extended mode). I just tested scrolling in this thread as well as changing between tabs, and I didn't have any problems. I ran with both Direct3D11 in Windowed and Exclusive mode without issue.

What are your settings? For reference I am currently running with:
Image Enhancement - FineSharp checked with settings of Strength 0.1 and thinning 0.035
Both of my upscaling are set to bicubic
Downscaling is Lanczos
Image doubling is super-xbr
Upscaling refinement is all unchecked
general settings - all checked except for delay playback start until render queue is full
ordered dithering
trade quality for performance - only thing checked is don't use linear light for dithering

If you have anything enabled that I mentioned as disabled on my system that may be your culprit. (Edit: You're using version 88.12, correct?)

[Magik Mark]

Setting madVR to PC levels (0-255) will avoid clipping. I've read in the JRiver Media Center guide that the graphics card should be set to TV levels, not the video renderer, if the TV expects 16-235.

The only way to test this theory is to run test patterns. My system consists with htpc, lumagen radiance mini & Pioneer Kuro (36bit) isf calibrated. My video processor though is set to pc levels

[huhn]

The only way to test this theory is to run test patterns. My system consists with htpc, lumagen radiance mini & Pioneer Kuro (36bit) isf calibrated. My video processor though is set to pc levels

madshi said madVR at 16-235 and GPU at 0-255 is better for a display that needs 16-235 in term of picture quality:
you can read it here:
http://www.avsforum.com/forum/139-display-calibration/1471169-madvr-argyllcms.html#post23274961

[XMonarchY]

So... what's a good FineSharp setting for HQ content and what's a good FineSharp setting for LQ content? I see some disagreement, but I definitely want the setting that does create (much) additional ringing and know whether to enable FineSharp in Image Enhancement and/or in Upscaling Refinement or both..?

[fairchild]

So... what's a good FineSharp setting for HQ content and what's a good FineSharp setting for LQ content? I see some disagreement, but I definitely want the setting that does create (much) additional ringing and know whether to enable FineSharp in Image Enhancement and/or in Upscaling Refinement or both..?

You could setup profiles for SD and HD like so. Personally I just use FineSharp in image enhancement for HD material with a strength of 1.5, and use only super-xbr for image doubling on SD with exact 2.0x scaling factor. So basically my image doubling only activates on content that is 960x540 or less (half of my screen resolution of 1920x1080)

if (srcHeight <= 540) and (srcWidth <= 960) "SD"
else if (srcHeight >= 541) and (srcWidth >= 961) "HD"

FineSharp - image enhancements OFF (SD)
FineSharp - image enhancements ON (HD)

[XMonarchY]

You could setup profiles for SD and HD like so. Personally I just use FineSharp in image enhancement for HD material with a strength of 1.5, and use only super-xbr for image doubling on SD with exact 2.0x scaling factor. So basically my image doubling only activates on content that is 960x540 or less (half of my screen resolution of 1920x1080)



FineSharp - image enhancements OFF (SD)
FineSharp - image enhancements ON (HD)

I used NNEDI3 for Chroma Upscaling (+ SuperRes filter)and Image Doubling/Quadrupling... Also, I saw some info on how Linear Light can create ringing. Is that something I should disable in Image Downscaling?

Does LumaSharpen add even more ringing and artifacts then FineSharp with NNEDI3 + SuperRest filter for Chroma Upscaling and Image Doubling/Quadrupling? In video games, LumaSharpen looks better than FineSharp, but I am uncertain about video playback. I've yet to see a comparison between FineSharp and LumaSharpen as far as overall resulting images and artifacts/ringing/aliasing goes. I mean I can capture an image myself and check it out, but I am not an expert on which shots to take for a good comparison.

[omarank]

Hmmmm... I've done some checks with Anime content now, and I have to say, although I was able to find some examples where "LL on" looked better, it was much easier to find examples where "LL on" screwed things up quite noticably. The BT.709 test build seems to sit in the middle, but from a quick check with Anime content, my vote right now would still go to "LL off", because there are more situations where "LL off" looks better compared to even the BT.709 middle ground test build.

Can anybody else besides iSunrise, aufkrawall and me go back to v0.88.11 and compare FineSharp with LL on/off with Anime content (and maybe double check with the test build)? There were several votes in favor of "LL on", so I don't want to revert that decision without some more "support". Thanks.
FineSharp LL: Despite the artifacts produced by LL ON setting, I find it makes the image more pleasing as compared to LL OFF setting. This may be due to the overall brightness of the image being closer to the original image when LL ON setting is used. I think that instead of considering LL OFF setting as a good compromise for all kinds of content, you should rather tune the LL ON setting such that the artifacts it produces can be removed/ reduced. As you will be investigating this possibility as per Shiandow’s suggestion, I am looking forward to seeing the tuned LL ON setting.

super-xbr: With soft source content, I like super-xbr for chroma upscaling. For image doubling, NNEDI3 looks better, but this is a good alternative for cases where NNEDI3 can’t be handled by the system.

NEDI chroma: I didn’t like it much. Jinc normally looks the best for chroma upscaling.

I haven’t tested super-xbr and NEDI chroma with SuperRes as Shiandow is already working to improve it. I will wait for him to release the improved version.

[madshi]

Some Finesharp settings i found i like on good old live action movies. (for a change :P )

- suggested strength for image enhancement (depends on grain and source quality, 1080p, no upscaling) - 0.4-0.6 low, 0.8-1.0 medium, 1.4-1.6 high (>2.0 ultra)
* thinning - 0.010 - 0.020 seems ok

- strength for upscaling refinement (720p, s-xbr doubling) - 0.6-0.8
* thinning - 0.000 - 0.005 - any noise or blocking is boosted by a large amount even with the lowest settings

- strength for upscaling refinement (SD, s-xbr quading) - <=0.4
* thinning - 0.000 - same as above and even more pronounced

s-xbr observations on same content

- used NNEDI32/64 for image doubling before trying s-xbr, can't find any reason to go back to NNEDI -_- . s-xbr is faster, sharper and i really can't find major drawbacks for image doubling use.

- tried it for chroma uspcaling on SD content and compared to NNEDI64 s-xbr is worse/different??? It's sharper, but there is some strange detail loss in some places.
Thanks for your feedback. I did find a bug in the super-xbr chroma upscaling algorithm, which will be fixed in the next build. Maybe that will solve the issue with the strange detail loss?

With D3D11 I do have lower rendering and presentation times. Is this a real performance gain or some pseudo effect like things happening behind madVRs back?

Edit: Clock state stays the same as with D3D9.
Probably using D3D9 and activating the "use a separate device for presentation" would give you the same benefit. At least that's my guess. D3D11 as used by madVR should not give you a performance benefit, except in very specific situations (in FSE mode if "present a frame for every VSync" is activated, and if your display refresh rate is higher than the movie frame rate).

At least I've managed to prevent SuperRes from introducing any aliasing
How? :)

I've make some tries today, and the repair function clearly seems to improve the aliasing control introduced by xstr.
So, i'll keep digging that way.
I've tried today to find a repair setting which works for every xstr value. Unfortunately it seems that if I set repair so high that the aliasing introduced by xstr is fixed, xstr doesn't have any benefit, anymore. So my latest impression is that xstr=thinning may have to be completely disabled... :(

The problem could probably be solved (or at least reduced) by using a large supersampling factor. But that would be too slow to make sense, I think.

quick question about image doubling. When you select super xbr or nedi, it becomes selected automatically for chroma, but when I check the stats with ctrl+j, it only shows super xbr/nedi for luma though. Does it only work on luma?
That sounds weird. Screenshot of the OSD?

additional question, why this rule does not working?
if (targetWidth != srcWidth) "A"
else "B"
I'm not sure. If this is a reproduceable problem, then please add it to the bug tracker - thanks!

I have found out what happens. It's pretty simple but equally annoying. My AMD graphics card switches on and off "High Performance Mode"
See? The old statistics simply averaged in this situation, leaving you with useless numbers.

But honestly I find very optimistic to think you can set 3 presets for sharpness... My TV has 4 options for sharpness, 2 of them have 100 levels. I agree that there is too many options atm but to reduce that at only 3 presets is maybe a little bit too extreme. What about just a strength setting for each sharpness filter? For Fine Sharp, you could for example set the strength and the thinning for each level. I don't say we have to find 100 levels of course but it will be a more flexible solution I think. You only have one parameter that controls everything but remains very flexible like you find on a TV.
I'm not sure where I'll end up. For now I want to reduce all of the "new" options (introduced in v0.88.0+) as much as possible. If reducing something won't make sense, then I won't do it. But I'm willing to sacrifice *some* flexibility to get simpler controls.

Judging by the current feedback, I may have to offer a slider for FineSharp instead of low/medium/high. But it should be just one slider, not two. And I'd still prefer to have low/medium/high. But we'll see...

I have been waiting for someone to report what I'm seeing but I guess I'm the only one. Since the introduction of DX11 version, whenever I choose it I get flickering video 30 second into the playback. And it's not disappearing until I restart the player or revert back to DX9. By flickering I mean alternating video frames with black frames. Like suddenly the video goes to 3D mode and shutter glasses activates over the screen. Please, tell me if this issue has been encountered already. I have a Radeon HD7990 with 15.5 Beta Drivers (Win 7 x64).
Do you have a similar effect when using D3D9 with the "use a separate device for presentation" enabled?

Do you have custom flush settings? Have you tried resetting madVR to default settings, and just enable D3D11?

Madshi, switching from FSE to windowed then back to FSE via right click is causing the render and present queues to struggle to stay full and I continue to get dropped frames.
This only happens when I have madVR set to display out at 10 bit with D3D 11, enabling present a frame for every vsync doesn't affect it at all.
But as long as you don't switch, the queues all fill just fine in both windowed and FSE modes? What happens if you disable the display mode switcher and set your display to 60Hz? Does that change anything? What happens if you lower the number of prepresented frames to 6? Does that "solve" the problem?

There is a queue problem in secondary monitor when there is a youtube (or other?) video in the primary monitor.
Do the queues fill back up automatically when you close the browser?

Yes, I noticed this too, although not the correlation with doing stuff on the primary monitor (I didn't look too close though). I noticed that with DX11 mode, I would get bunches of dropped frames periodically (like, a bunch in a row), and sometimes the queues would fail to fill until I paused and unpaused. With DX9, the presentation glitches counter would steadily rise (causing visual glitches), but it wouldn't actually drop any frames on madVR's end.

I have no idea whether or not this is a regression though - the presentation glitches seem like a DWM problem, but perhaps the DX11 behavior could be improved.
If you can find a reproducable case which does *not* consist of some other software also using the GPU, please let me know.

FineSharp LL: Despite the artifacts produced by LL ON setting, I find it makes the image more pleasing as compared to LL OFF setting. This may be due to the overall brightness of the image being closer to the original image when LL ON setting is used. I think that instead of considering LL OFF setting as a good compromise for all kinds of content, you should rather tune the LL ON setting such that the artifacts it produces can be removed/ reduced. As you will be investigating this possibility as per Shiandow’s suggestion, I am looking forward to seeing the tuned LL ON setting.
Unfortunately I've tried and failed, at least for now. I haven't managed to make FineSharp maintain the overall brightness without introducing noticeable artifacts. Please try some Anime content and watch black/dark lines. Compare with v0.88.11, so you can turn LL on/off. I think you will have no problem finding several instances where LL on produces significant artifacts. Because of that I've now disabled LL for now in the latest build. Maybe I'll find a solution at some time in the future. But for now I consider the artifacts too strong and they occur too often to keep using LL on.

super-xbr: With soft source content, I like super-xbr for chroma upscaling. For image doubling, NNEDI3 looks better, but this is a good alternative for cases where NNEDI3 can’t be handled by the system.

NEDI chroma: I didn’t like it much. Jinc normally looks the best for chroma upscaling.
Thanks for the feedback. I don't like NEDI for chroma upscaling, either. FWIW, I've found a little bug in super-xbr chroma upscaling. Please compare again. Maybe super-xbr could replace Jinc as the "best medium speed" chroma upscaler?

[madshi]

madVR v0.88.13 released

http://madshi.net/madVR.zip

* added super-xbr "sharpness" parameter (50, 75, 100 = default, 125)
* big speed for super-xbr chroma upscaling
* small speedup for super-xbr image doubling
* FineSharp no longer uses linear light (for now)
* added IMadVRCommand interface, replaces IMadVRSeekbar/ExclusiveModeControl
* added IMadVRCommand::SendCommand("restoreDisplayModeNow") command
* fixed: small logic/quality bug in super-xbr chroma upscaling
* fixed: settings dialog always showed NNEDI3 chroma doubling to be enabled
super-xbr chroma upscaling is more than twice as fast as before with my AMD HD7770 now. It's now just moderately slower than JincAR. Just chroma upscaling got that much faster, though, not luma/chroma/image doubling.

Still looking for more FineSharp feedback. Seperately for image enhancements and upscaling refinement. Thanks.

[Eyldebrandt]

I've tried today to find a repair setting which works for every xstr value. Unfortunately it seems that if I set repair so high that the aliasing introduced by xstr is fixed, xstr doesn't have any benefit, anymore. So my latest impression is that xstr=thinning may have to be completely disabled... :(

The problem could probably be solved (or at least reduced) by using a large supersampling factor. But that would be too slow to make sense, I think.


That was exactly my final opinion of FineSharp in avisynth.
And by supersampling, you will also affect the efficiency of the sharpen .
My personal though about FineSharp, w/e the version, avisynth or madVR, is that it's not a good sharpen for HQ content. And he's absolutely horrible with SD content.

In years, LSF still seems to me the best sharpener for video content in real time.
Even further, LSFHC is the best for HQ content.

Other thing, it appears on some tests i made than SuperRes (in upscaling refinement) seems to be extremely destructive in some cases, even with really moderate settings.
I have some demonstratives screenshots but as Shiandow is currently working on SuperRes, maybe it's not the time to post them.
Let me know.

[tFWo]

Thanks for your feedback. I did find a bug in the super-xbr chroma upscaling algorithm, which will be fixed in the next build. Maybe that will solve the issue with the strange detail loss?

I have no idea if it's s-xbr detail loss or nnedi3 artifacts, so here is a comparison. The difference is small and with the new s-xbr speedup i'll just leave s-xbr100 on for chroma upscaling :D .

nnedi364+arrow
http://i.imgur.com/Za4Hk6N.png
nnedi364
http://i.imgur.com/ZndAo5P.png
s-xbr100
http://i.imgur.com/FbLKWXL.png
s-xbr125
http://i.imgur.com/XK11P6x.png

[Anima123]

madVR v0.88.13 released

http://madshi.net/madVR.zip

* added super-xbr "sharpness" parameter (50, 75, 100 = default, 125)


madshi, does the "sharpness" 100 or 125 is equivalent to the previous default version, which is Sinc?

And is the"sharpness" parameter the same as "filter weights" Hyllian was talking about?

[Ver Greeneyes]

If you can find a reproducable case which does *not* consist of some other software also using the GPU, please let me know.I haven't had any time to do testing recently, but I'll let you know when I've tried it again.

[Hyllian]

madshi, does the "sharpness" 100 or 125 is equivalent to the previous default version, which is Sinc?

And is the"sharpness" parameter the same as "filter weights" Hyllian was talking about?

The question wasn't for me, but I can help here.

The answer is yes for both questions! The sharpness 100 is a sinc equivalent value and the sharpness params control the filter weights. When you go down on these values, you aproximate to cubic coefficients around 50 (0.0 would mean pure bilinear). Above 100 is like "above speed light". :D

[har3inger]

Reporting back on very early first impressions for super-xbr.

super-xbr for image upscaling seems noticeably faster (thank you!) and chroma super-xbr is really, really fast compared to how it was before. I'm trying it for chroma now instead of jinc3 ar because I'm finding that it does good things in image upscaling without obvious strange artifacts.

For DVD content where you quadruple image res (720x384-> ~1080p), if you set the double to 75 sharpness and the quadruple to 50 sharpness, you get a very similar sharpness to jinc3 ar but with *way* less aliasing. I'm currently trying out 50 sharpness/50 sharpness for double/quad super-xbr for now, even though it's softer than jinc3ar, because it's much cleaner (less aliasing, less source artifacts, but mediocre sharpness). At these lower sharpness values (50 or 75), haloing is not an issue. Source ringing is reduced, in fact. I cannot compare to nnedi3, because it doesn't work for me, but I imagine this must be close to how nnedi3 users must see things. For these tests, I didn't use SuperRes because in its current form, I think it magnifies source artifacts too much for these tiny source resolutions.

For 720p content, I've cranked up sharpness for image doubling to 125, and will test drive it like this for a while just for laughs. So far, it has actually been alright when paired with SuperRes (1 pass, 0.75 str, 0.2 sharp, 0.1 soft, 0 AA, 0AR) as long as the source is very clean. I suspect I might not keep the setting that high for long though.

[Hyllian]

I'm currently trying out 50 sharpness/50 sharpness for double/quad super-xbr for now, even though it's softer than jinc3ar...

Now that you mentioned, it reminded of something. Does madVR allow using two different shaders in a same task? Let's say you want to upscale by 4x, does it allow upscale the first 2x using Super-xBR and the last 2x using Jinc3AR? (I'm not a video player user.)

That's because in my Retroarch use, I find it much better using Super-xBR only for the first 2x upscaling and a Jinc for the last 2x. The sharpness is much better than using Super-xBr two times. I think the first 2x filter imprints its characteristics to the final output. And the second one does a much easier work to scale from 2x to 4x, as the aliasing was already fixed by the first one. And as jinc is a very good linear filter, it does a very good job as a second pass filter.

[har3inger]

Yes, sort of. You just turn off letting Madvr image quadrupling (uncheck the box) and it'll use whatever you have for "image upscaling" to complete the upsize if the initial doubling is insufficient to hit target resolution. You however cannot specify two different non-doubling algorithms (like jinc3ar followed by bicubic75ar) or two different doubling algorithms (first nnedi3, then super-xbr) at this time from the UI, but I think there'd be very little reason to need those sorts of setups.

Edit: I lied, you can totally set up two different doubling algorithms to work in sequence. (eg first nnedi3, then super-xbr). And now that I go back to test it, I'm having a very difficult time telling the difference between > super-xbr50 > super-xbr50 < catmull-rom AR and > super-xbr50 > Jinc 3 ar (Also, I find those upscale/downscale triangle brackets super unintuitive. < to me means upscale because it starts small then gets bigger as you read left to right)

[JonnyRedHed]

Is the image upscale super-xbr supposed to be like this. Unselectable for all the image upscale sections I have.

http://i62.tinypic.com/2h7ie5w.png

[har3inger]

Super-xbr should only be selectable under image doubling and chroma upscaling. I'm not sure why that entry is even there for you under image upscaling. It's not for me.

[Warner306]

Is the image upscale super-xbr supposed to be like this. Unselectable for all the image upscale sections I have.

http://i62.tinypic.com/2h7ie5w.png

I have the same problem. This entry shows up under existing profiles.

[mindz]

For me aswell.

[madshi]

That was exactly my final opinion of FineSharp in avisynth.
And by supersampling, you will also affect the efficiency of the sharpen .
My personal though about FineSharp, w/e the version, avisynth or madVR, is that it's not a good sharpen for HQ content. And he's absolutely horrible with SD content.

In years, LSF still seems to me the best sharpener for video content in real time.
Even further, LSFHC is the best for HQ content.

Other thing, it appears on some tests i made than SuperRes (in upscaling refinement) seems to be extremely destructive in some cases, even with really moderate settings.
I have some demonstratives screenshots but as Shiandow is currently working on SuperRes, maybe it's not the time to post them.
Let me know.
FWIW, Shiandow might be interested in samples of files which the old/current SuperRes struggles with - to make sure the new algo handles them better.

I have no idea if it's s-xbr detail loss or nnedi3 artifacts, so here is a comparison. The difference is small and with the new s-xbr speedup i'll just leave s-xbr100 on for chroma upscaling :D .

nnedi364+arrow
http://i.imgur.com/Za4Hk6N.png
nnedi364
http://i.imgur.com/ZndAo5P.png
s-xbr100
http://i.imgur.com/FbLKWXL.png
s-xbr125
http://i.imgur.com/XK11P6x.png
Not sure myself which is more "accurate".

Reporting back on very early first impressions for super-xbr.

super-xbr for image upscaling seems noticeably faster (thank you!) and chroma super-xbr is really, really fast compared to how it was before. I'm trying it for chroma now instead of jinc3 ar because I'm finding that it does good things in image upscaling without obvious strange artifacts.

For DVD content where you quadruple image res (720x384-> ~1080p), if you set the double to 75 sharpness and the quadruple to 50 sharpness, you get a very similar sharpness to jinc3 ar but with *way* less aliasing. I'm currently trying out 50 sharpness/50 sharpness for double/quad super-xbr for now, even though it's softer than jinc3ar, because it's much cleaner (less aliasing, less source artifacts, but mediocre sharpness). At these lower sharpness values (50 or 75), haloing is not an issue. Source ringing is reduced, in fact. I cannot compare to nnedi3, because it doesn't work for me, but I imagine this must be close to how nnedi3 users must see things. For these tests, I didn't use SuperRes because in its current form, I think it magnifies source artifacts too much for these tiny source resolutions.

For 720p content, I've cranked up sharpness for image doubling to 125, and will test drive it like this for a while just for laughs. So far, it has actually been alright when paired with SuperRes (1 pass, 0.75 str, 0.2 sharp, 0.1 soft, 0 AA, 0AR) as long as the source is very clean. I suspect I might not keep the setting that high for long though.
Glad to hear you like it!

Now that you mentioned, it reminded of something. Does madVR allow using two different shaders in a same task? Let's say you want to upscale by 4x, does it allow upscale the first 2x using Super-xBR and the last 2x using Jinc3AR? (I'm not a video player user.)

That's because in my Retroarch use, I find it much better using Super-xBR only for the first 2x upscaling and a Jinc for the last 2x. The sharpness is much better than using Super-xBr two times. I think the first 2x filter imprints its characteristics to the final output. And the second one does a much easier work to scale from 2x to 4x, as the aliasing was already fixed by the first one. And as jinc is a very good linear filter, it does a very good job as a second pass filter.
For my taste your Jinc implementation rings too much. In all my tests I preferred running super-xbr twice instead of following up super-xbr with Jinc.

Anyway, users do have the choice. They can use super-xbr once or twice (with same or different sharpness settings), or even multiple times, and they can use any linear algorithm afterwards to get to the final target resolution.

Is the image upscale super-xbr supposed to be like this. Unselectable for all the image upscale sections I have.
That option is not supposed to be even visible. The reason it's visible is probably because you've set your OS to higher DPI settings, or something like that, which seems to be screwing with the layout of my settings dialog. The settings page for "image upscaling" is supposed to be sized so that the "super-xbr" checkbox isn't visible. I'm reusing the same settings page for all the image/chroma up/downscaling options, and just hide some settings or resize things or move things around a little.

Anyway, super-xbr is not available for image upscaling, anyway, because it can only exactly double resolution. So this is just a cosmetical issue, nothing else.

[Skwelcha]

Found a more or less cosmetical bug as well. If you choose s-xbr in chroma upscaling and choose bicubic right after that, the drop down menu for the sharpness seems to be the one from s-xbr, because you can go up to 125.

[JonnyRedHed]

That option is not supposed to be even visible. The reason it's visible is probably because you've set your OS to higher DPI settings, or something like that, which seems to be screwing with the layout of my settings dialog. The settings page for "image upscaling" is supposed to be sized so that the "super-xbr" checkbox isn't visible. I'm reusing the same settings page for all the image/chroma up/downscaling options, and just hide some settings or resize things or move things around a little.

Anyway, super-xbr is not available for image upscaling, anyway, because it can only exactly double resolution. So this is just a cosmetical issue, nothing else.


I see, thanks for explaining. Yes my windows is set to 125% DPI scale. Am using a large plasma TV for windows, and my eyes struggle to read at 100%.


Also I was wondering if there could be a way to add in new lines (features) to existing profiles. Like in my screen grab below the 'upscaling refinement' sits above my 4 profiles from when it was introduced some time back. But I would like to add it into all four of my profiles, can this be done?


http://i62.tinypic.com/2h7ie5w.png

[Shiandow]

At least I've managed to prevent SuperRes from introducing any aliasing, so when combined with super-xbr or NEDI you shouldn't need anti-aliasing, which didn't work too well anyway.

How? :)


I've experimented a bit with the "bilateral backprojection" method described in this paper (http://www.ece.northwestern.edu/~yingwu/papers/conference/2007/ICME07_Dai_final.pdf). Although I've made some changes to it, making it more similar to the "Joint bilateral upsampling" method described in this paper (http://www.johanneskopf.de/publications/jbu/index.html). All in all these changes prevent the algorithm from introducing aliasing and ringing, although it will still try to reproduce them if they are present in the original image.

FWIW, Shiandow might be interested in samples of files which the old/current SuperRes struggles with - to make sure the new algo handles them better.

That would be handy, although I'd prefer samples were the new algorithm performs badly. The new algorithm is a whole different beat with it's own weak points, and as far as I know the only thing both algorithms can't handle well is bad sources (high ringing / aliasing), in those cases NNEDI3 tends to work better since it couldn't reproduce aliasing and ringing even if it wanted to.

[madshi]

Found a more or less cosmetical bug as well. If you choose s-xbr in chroma upscaling and choose bicubic right after that, the drop down menu for the sharpness seems to be the one from s-xbr, because you can go up to 125.
Yep, another bug.

Also I was wondering if there could be a way to add in new lines (features) to existing profiles. Like in my screen grab below the 'upscaling refinement' sits above my 4 profiles from when it was introduced some time back. But I would like to add it into all four of my profiles, can this be done?
Currently you have to delete the profiles and redo them. Of course it would technically possible for me to add a feature to extend profiles like this in a future madVR version, it would cost me too much time at this point. Maybe in a far away future version, but not soon.

I've experimented a bit with the "bilateral backprojection" method described in this paper (http://www.ece.northwestern.edu/~yingwu/papers/conference/2007/ICME07_Dai_final.pdf). Although I've made some changes to it, making it more similar to the "Joint bilateral upsampling" method described in this paper (http://www.johanneskopf.de/publications/jbu/index.html). All in all these changes prevent the algorithm from introducing aliasing and ringing, although it will still try to reproduce them if they are present in the original image.
I had the 2nd paper in my list of papers to look at, in the hope that it might allow us to use the luma information to guide in chroma upsampling. Is that something that sounds like worth investigating to you?

[James Freeman]

madshi, there are bad queue bugs in the 88.12 and 88.13 builds when using dual monitor setup, DX9 and DX11.

Just go fullscreen (windowed) in the secondary monitor, scroll the browser (make sure the browser window is in focus) in the primary monitor up and down like crazy with the mouse wheel and look at the queues in the secondary monitor.
Or move the browser window in the primary monitor from left to right frantically.
This is the fast way to recreate this bug. Normally it happens by itself on my system.

EDIT:
This happen in lower Power State of the GPU when the frame rendering times go higher than 42ms.
I may be just wasting your time... Sorry about that.

[JarrettH]

Source: H.264 8bit 4:2:0

1. chroma upscaling: JincAR3
2. chroma upscaling: NNEDI16
3. chroma upscaling: Super-XBR
4. chroma upscaling: JincAR3 with SuperRes
5. chroma upscaling: NNEDI16 with SuperRes
6. chroma upscaling: Super-XBR with SuperRes

SuperRes setting:
passes: 2
strength: 0.75
sharpness: 0.25
softness: 0.10
anti-aliasing: 0.50
anti-ringing: 1.00

1 4
2 5
3 6
http://s5d4.turboimagehost.com/t/23368978_up.png (http://www.turboimagehost.com/p/23368978/up.png.html)

http://s5d4.turboimagehost.com/t/23368979_upx3.png (http://www.turboimagehost.com/p/23368979/upx3.png.html)



One other thing: did you notice those blue artifacts in #3?

[har3inger]

Hmm, yeah, those artifacts are pretty bad (blue blotching/ringing inside the white japanese characters of the zoomed speech bubbles). They're there in jinc3 (image 1) as well, but not as prominent. In my testing with the latest build chroma super-xbr does indeed emphasize color ringing/noise artifacts if they're in the source (I check using nearest neighbor and bilinear). That comparison might just be showing off nnedi3's ability to reduce source ringing artifacts, as we've already seen with other examples.


As an aside, has anyone tried using the current superres with 0 set for sharpness, softness, AA and AR on super-xbr luma? I'm finding that it's actually quite good across all types of sources if you tweak the strength/passes appropriately.

[Anima123]

As an aside, has anyone tried using the current superres with 0 set for sharpness, softness, AA and AR on super-xbr luma? I'm finding that it's actually quite good across all types of sources if you tweak the strength/passes appropriately.

I like the effect of SuperRes when sharpness & softness be set to 0, but AA and AR been kept, for 576p videos mostly.

It's bad for lower resolution videos though, I would like to try setting AA & AR to 0 for those files and see if it's more appropriate.

[leeperry]

Oh wow, :thanks: for the new build with the s-xbr sharpness setting, I still seem to prefer NEDI for 720p@1080p because it looks more natural with less EE, but on noisy 320*240 s-xbr@50 seems to extract quite a bit more details than NNEDI3 huh. Looks like I won't be tied to the 13.12 drivers forever after all :p

Of course I use SR on top of all of them and I finetune its sharpness settting each time. I really like to rely on SR's sharpness setting and even though s-xbr@50 looks very promising from my limited testing so far, would there be any way to go lower? Or possibly allow us to choose the sharpness figure please? :o

[JarrettH]

I think superres is very nice. It doesn't feel like a post-process - the others seem to bring out about as much bad as good, as can be seen by some preferring a setting as closest to off as possible lol

My superres setting is

Two passes, 720p to 1080p
0.45, 0.25, 0.20, 0.20, 0.85

[madshi]

One other thing: did you notice those blue artifacts in #3?
Hmmmm... I didn't notice them until now. I think they might be caused by the strict AR filter that I'm using for super-xbr chroma upscaling. Basically I'm using the strict filter from the original shaders, because I thought it wouldn't be too important for chroma upscaling. Maybe I have to replace that with my own, too. But it's hard to say if that would help without having access to a sample where the problem shows up.

@viewer, can you provide a small sample of the source?

Oh wow, :thanks: for the new build with the s-xbr sharpness setting, I still seem to prefer NEDI for 720p@1080p because it looks more natural with less EE, but on noisy 320*240 s-xbr@50 seems to extract quite a bit more details than NNEDI3 huh. Looks like I won't be tied to the 13.12 drivers forever after all :p

Of course I use SR on top of all of them and I finetune its sharpness settting each time. I really like to rely on SR's sharpness setting and even though s-xbr@50 looks very promising from my limited testing so far, would there be any way to go lower? Or possibly allow us to choose the sharpness figure please? :o
No free sharpness slider or dial planned right now. Why would you want to go even softer than 50? It's the softest I would go for super-xbr. If you go much softer, at some point you're going to lose the ability to recover the image detail through sharpening/SuperRes. I mean, if you take softening to the extreme, you would end up with just a gray (or slightly colored) frame. Even sharpening/SuperRes can't get something useful out of that.

-------

No real benefit discussing SuperRes right now because it will change soon. And probably more than just a small bit, I suspect. But we'll see...

[Shiandow]

I had the 2nd paper in my list of papers to look at, in the hope that it might allow us to use the luma information to guide in chroma upsampling. Is that something that sounds like worth investigating to you?

I've actually tried that already, in fact I've added it as an option for the new SuperChromaRes (although it can be used separately if you want). Interestingly you can do the same thing in the upscaling step for the SuperChromaRes algorithm. Even though you could also use the chroma information at that point, doing so tends to make the image worse. I'm not too sure why this happens, but by only using the luma information you force the chroma and luma to match, and you can be sure that the luma is correct which might also improve image quality.

[viewer]

@viewer, can you provide a small sample of the source?

sample here:
https://mega.co.nz/#!bR8mjSCT!7ki_LvGAtfl_VEV73b_Q5iNRZh-IZAMxXB1mYPdTxvs

[SecurityBunny]

Madshi, I have another case of queues not filling all the way when toggling between fullscreen, windowed, and fullscreen again. It is the render queue that drops (0/1-2 from 7-8), which subsequently drops the present queue fill rate. Just noticed since updating. I have to pause playback for a second for queues to fill all the way up again.

Going into fullscreen initially, queues are fine. The issue occurs when going fullscreen, then windowed, then back to fullscreen again. Please let me know if you need me to debug with a special file to help narrow down where the problem may be.

D3D11 10-bit
MPC-HC 1.7.9.25 x64
MadVR 0.88.13
Nvidia GeForce 353.30

[James Freeman]

Just to add to that queue problem,

In dx11 mode if the rendering time goes above frame time (ie. 42ms in 24p) the queues will empty (as expected), BUT, they will not fill back up when the rendering time goes down to below frame time and stay at 1-2/8.
In dx9 (new path) they fill back as normal.
When AERO is off and dx9 old path is active I think there is a queue problem also.
I'll check again.

[vomanci]

Possible bug : in windows 8.1, if i chose different scaling (DPI) for my monitor and TV, MadVR doesn't enter fullscreen, windowed or FSE, on my TV with DX11 path. It changes from windowed to fullscreen but the OSD still shows windowed mode (no fullscreen). And no search bar at the bottom.

[nevcairiel]

Possible bug : in windows 8.1, if i chose different scaling (DPI) for my monitor and TV, MadVR doesn't enter fullscreen, windowed or FSE, on my TV with DX11 path. It changes from windowed to fullscreen but the OSD still shows windowed mode (no fullscreen). And no search bar at the bottom.

Thats probably your media player not supporting different DPI modes. That happens with MPC-HC for me when I have one screen at high-dpi (ie. my 4K screen), and all other screens on normal DPI.

[vomanci]

Thanks. It's MPC-HC.

[ryrynz]

But as long as you don't switch, the queues all fill just fine in both windowed and FSE modes? What happens if you disable the display mode switcher and set your display to 60Hz? Does that change anything? What happens if you lower the number of prepresented frames to 6? Does that "solve" the problem?


Yes, display mode switcher is currently off and refresh is already 60Hz.

Changing prepresented frames for Exclusive mode has no effect unless I set it to 1. If I set it to 2+ I get glitches on the screen (part flashes of previous frames) even though no glitches or dropped frames show on the OSD.
If I change prepresented frames to 1, everything seems to work fine.

[leeperry]

No free sharpness slider or dial planned right now. Why would you want to go even softer than 50? It's the softest I would go for super-xbr. If you go much softer, at some point you're going to lose the ability to recover the image detail through sharpening/SuperRes.
Because there's only one way to find out? 50 works very nicely on hopeless tiny videos but it's way too sharp for 720p@1080p IME, I'm not seeking hardcore details extraction in that latter case as I just want something sharper than J3AR...which I do get with NEDI, but I would love to try a softer x-sbr if any possible. It might very well be in order to compensate SR's sharpness setting but at the end of the day all roads lead to Rome and I really like SR so my goal is mostly to try all possible options and take it from there :)

[leeperry]

Yeah, I just ran a few more tests on 576*320@1080p and s-xbr@50 is much sharper than NNEDI3 IME, if going any lower is out of the question could I use it on my own via a PS script in my media player please? 50 is prolly very nice without SR and soft scaling but I would like to keep my other settings as is because they look great, I just need s-xbr to come in less strong :o

[Ge'in]

super_xbr at 100 of sharpness, it's just perfect for me. For chroma upscaling and luma doubling whatever the video source (sd or hd).
It's a very good compromise between quality/speed.

[mindbomb]

That sounds weird. Screenshot of the OSD?


What I meant was that when using super xbr, the osd reports "chroma>Lanczos 3 AR, luma>super xbr <catmull rom" but it would be more descriptive if it were "chroma>super xbr>lanczos, luma>super xbr <catmull rom", since super xbr does work on both chroma and luma, right?

[michkrol]

Are you sure it doesn't say image > super xbr?
Image means madVR scales "whole" image, both chroma and luma at the same time, that's why OSD doesn't report chroma and luma processing separately.
Please post a screenshot as madshi has asked you to.

It does work correctly for me, first upscaling chroma, then doubling both chroma and luma (i.e. image), then upscaling with Jinc to target resolution.
http://i.imgur.com/PjAQu1B.png

[mindbomb]

oh, okay, you are right michkrol, i see how it works now.

[VenomousNinja]

Activating use of Direct3D 11 in MadVR on my setup makes my secondary monitor connected via HDMI (LG TV) to do weird things during avi payback. GTX970, Win 7 x64, latest driver 353.30, MPC-BE. Exclusive full screen. Monitor just goes black 2-3 times during video playback for a second or two.I'd say every 20-25 mins. Sometimes it's just had been stuck in some kind of signal loss loop - black-video-black-video. Even after closing video player (MPC-BE) it's still remains in on - off loop. I've deselected dx11 and usinf DX9 for now. No mystic happens. This kind of stuff started to happen in the last 2-3 build of MadVR. Possibly it's nvidia new drivers...

[XMonarchY]

I tested FineSharp on several sources and each time I could see obvious ringing using FineSharp in image enhancements and upscaling refinement. LumaSharpen does not SEEM to do that at default settings, so it is beyond me how people find FineSharp superior to LumaSharpen...