madVR - high quality video renderer (GPU assisted)

[nevcairiel]

Although, I still don't like having downscaling under "image upscaling."

I think its perfectly logical - you want to upscale the image, and all options that affect this process should be combined in one place. It doesn't seem logical that the "downscaling" options would affect the upscaling process - no matter what it does internally.

Additionally, it allows you to use a different setting for actual downscaling without having to go to profiles, which I also like as a simplification.

[Mistery 73]

a greeting to all of you I need your advice on what video card to buy using madVR in 4k ... I was focused on this ZOTAC GeForce 1080 GTX AMP 8GB Extreme Edition .. accept any advice .. thanks

[ajp_anton]

It does in the OSD.


Again, the OSD is your friend, It changes things for me..
704x480 > 1920x1200

Normal Bicubic 60 > Bicubic 60
High NGU low > Bicubic 60
Very High NGU Med > NGU Low.


Do u even OSD?


Nobody else has cared about this, it's pretty self explanatory because there's no option for chroma.. and guess what..
If you use the OSD..

Chroma > DXVA
Yes, I already know the answers to all of those, by following this thread. Just saying that not all people have read all posts in this ginormous thread. And the OSD should *not* be the place to check what your settings do, it should not be guesswork. What if all image upscaling options were just "option 1", "option 2" etc, and you'd have to just try and see in the OSD what they are.

And no, the fact that the DXVA option doesn't exist for chroma doesn't really explain anything. In fact, someone might think it means that DXVA can't do chroma at all.

[ajp_anton]

it could be if you actually had the option to set the exact point of time at which a frame should be displayed, which isn't the case. that's why g/freesync so far is useless for videos.
I meant the variable rendering time in games. The lack of being able to display at exact times is an equally big problem for video and games, but games also has the additional disadvantage of having to display ASAP, no matter how quickly or slowly the image was rendered, adding additional jitter.

[ryrynz]

Yes, I already know the answers to all of those, by following this thread.


Yeah, pretty sure what you want isn't going to happen.

[Sp00kyFox]

I meant the variable rendering time in games. The lack of being able to display at exact times is an equally big problem for video and games, but games also has the additional disadvantage of having to display ASAP, no matter how quickly or slowly the image was rendered, adding additional jitter.
regarding gsync this is actually an advantage because you don't have the problem to display a frame at a certain point in time when you just push them out ASAP. the jitter introduced by variable render timings though is something you just have to live with in the field of video games. even gsync can't fix this. but the scene complexity of a virtual environment usually changes in a continuous way so does the render timing. so this is not a big deal as long as it doesn't lead to framerate drops in a gsync-free display setup.

[burfadel]

Problem with supporting variable frame rate is that there are two implementations. The first is an open standard by AMD and is actually part of the VESA spec. Cost of implementation is low because it is part of the standard anyway. The other one is closed sourced proprietary G-Sync by Nvidia, which requires an Nvidia card and a specialised G-Sync monitor that has the G-Sync module. A G-sync monitor can also be Freesync, the only limiting thing is Nvidia probably doesn't allow them? In any case, if VFR is supported it would make more sense to support Freesync.

[StinDaWg]

The biggest issue I have with the current way things are now with image doubling, is that there is no way to select which luma algorithm to use after doubling. It is locked to Bicubic 60AR, and that is unacceptable.

If I'm playing back a 720x404 video on a 1080p screen, the image is doubled to 1440x808 and then needs to be upscaled to 1920x1080. I want to use Jinc for that, but I can't. What logic is that? I suppose if NGU can eventually do 3x or 4x scaling, then all I would need to do is downscale after, but who knows how much more gpu that will use.

[burfadel]

The biggest issue I have with the current way things are now with image doubling, is that there is no way to select which luma algorithm to use after doubling. It is locked to Bicubic 60AR, and that is unacceptable.

If I'm playing back a 720x404 video on a 1080p screen, the image is doubled to 1440x808 and then needs to be upscaled to 1920x1080. I want to use Jinc for that, but I can't. What logic is that? I suppose if NGU can eventually do 3x or 4x scaling, then all I would need to do is downscale after, but who knows how much more gpu that will use.

I agree, it would be good to be able to select the final upscaler. Maybe a limited list can be provided, such that there is the default, Jinc etc. In MPDN player on here that is an all-in-one player, there are other options not seen in madVR. I wonder if they are any better. Such as sinc-Hann, sinc-Jinc, Sinc-Hamming, option for 8 taps etc. Maybe having the options for the best variation of these would be good :).

[Ver Greeneyes]

There you go.
Oops, didn't remember that well enough it seems :p

[thighhighs]

People talk about g-sync here, but what about another game technology - multi-gpu support? Because some users now have two gpu's or dual gpu like 295x2 and for madvr it completely useless or negative. I hope madvr can support SLI\CF in future.

[aufkrawall]

Drivers are buggy enough even without multi-GPU.

[thighhighs]

Drivers are buggy enough even without multi-GPU.
Why it should be buggy? SLI (i use nvidia) in old games work fine with new drivers. I dont think that possible problems worse vs todays useless second gpu and negative scaling. But if SLI can boost madvr performance it will be great.
edit: I read some posts about multi-gpu here, but i don't understand, madshi don't want introduce it or it just impossible.

[fedpul]

Hi guys. I have a challenge for you. You must download them and compare them in fullscreen and try to identify which Option goes with the right settings. After some time I will be posting the results.

This is the link: http://imgur.com/a/zmRYh

Movie Resolution: 720x480
Output Resolution: 1920x1080

In both options I just doubled resolution once and only in Luma, so downscaling algo is not relevant. So this test is basically to see if you are able to see differences after doubling between Jinc AR and Bicubic 60 AR.

One of the options is with 0.91.1. Settings:
Remove Banding: Medium/High
Chroma: NGU High
Luma Doubling: NGU High
Image Upscaling: Jinc AR
Downscaling: Not relevant.
Add grain and Soften Edges at 1

The other is with 0.91.3 version. Settings:
Remove Banding: Medium/High
Chroma: NGU Very High
Luma Doubling: NGU Very High (all auto)
Downscaling: Not relevant.
Add grain and Soften Edges at 1

[ashlar42]

I thought it was due to PotPlayer that I sometimes got a black screen for a second or two when closing it but creating a "madVR\DontRenderAfterStop" empty folder fixed it as far as I can tell. Prolly unrelated to your problem, though as I always was able to at least open the taskmanager and kill PotP.
What's this option? I've never seen it mentioned before.

[Backflash]

Hi guys. I have a challenge for you. You must download them and compare them in fullscreen and try to identify which Option goes with the right settings. After some time I will be posting the results.

This is the link: http://imgur.com/a/zmRYh


Are you trying to say option 2 is jinc?
though that aliasing on the hat, option 2 looks better to me.

[Dodgexander]

I tried with mpc and evr and it doesn't happen but I think that's because the PC just doesn't go to sleep when using evr.

With zoom player and MadVR I disabled the screen saver and power saving prevention and it doesn't sleep.

So why is mpc hc sleeping with paused playback with madVR and not with evr? Why does it bsod even if it does?

Sent from my Nexus 5 using Tapatalk

[pirlouy]

@thighhighs:SLI/crosfire is complicated for developers:
http://www.gamedev.net/topic/666419-what-are-your-opinions-on-dx12vulkanmantle/#entry5215019
Multi GPU (SLI/CrossfireX) is fucking complicated. You cannot begin to conceive of the number of failure cases that are involved until you see them in person

[fedpul]

Are you trying to say option 2 is jinc?
though that aliasing on the hat, option 2 looks better to me.
I won't tell you know. As I said I will post the results later. Just remember what options you pick.

Enviado desde mi MI 4LTE mediante Tapatalk

[omarank]

Hi guys. I have a challenge for you. You must download them and compare them in fullscreen and try to identify which Option goes with the right settings. After some time I will be posting the results.
Looks like the image in Option 1 (u2xyyDQ.png) is the one scaled with Jinc after NGU doubling. I believe though this kind of challenge would be more fair for movie content and not anime content.

[Xaurus]

Ok so I've jumped onto the latest madvr with this new NGU scaler (which is excellent), but I must have missed something about the change of all the settings. Where did image doubling go? I understand that the settings dialogues have been changed but I can't find any info from madshi in his latest updates in this thread.

My point is, I find the new settings kind of confusing, that's all. Especially with regards to the image doubling.

[fedpul]

Looks like the image in Option 1 (u2xyyDQ.png) is the one scaled with Jinc after NGU doubling. I believe though this kind of challenge would be more fair for movie content and not anime content.
This weekend I will be posting the results, and I will upload the same challenge but with a movie instead of anime, the next challenge will last one Week, from one weekend to another.

Enviado desde mi MI 4LTE mediante Tapatalk

[seiyafan]

Ok so I've jumped onto the latest madvr with this new NGU scaler (which is excellent), but I must have missed something about the change of all the settings. Where did image doubling go? I understand that the settings dialogues have been changed but I can't find any info from madshi in his latest updates in this thread.

My point is, I find the new settings kind of confusing, that's all. Especially with regards to the image doubling.

Doubling and upscaling have merged into one. So in upscaling if you choose NGU or NNEDI3 it's doubling.

[cork_OS]

Looks like the image in Option 1 (u2xyyDQ.png) is the one scaled with Jinc after NGU doubling. I believe though this kind of challenge would be more fair for movie content and not anime content.
Agreed, it's a bit more focused/sharp.

[kalston]

People talk about g-sync here, but what about another game technology - multi-gpu support? Because some users now have two gpu's or dual gpu like 295x2 and for madvr it completely useless or negative. I hope madvr can support SLI\CF in future.

Multi-GPU is a complicated mess and everyone is giving up on it for a reason - game devs and AMD/nvidia are all dropping the ball. Only DX12 could save it but that requires extra work (and training) for the game/app developers. I think only ONE game has dx12 multi GPU support as of right now (the last Deus EX IIRC? and that's not a really well optimized game to begin with).

Just look at how few recent releases support multi-GPU and how bad it is most of the time when they actually do. Poor scaling, bugs, (micro)stuttering etc...

This is even more ridiculous than asking to support VRR for video playback IMO.

[ROKUGAN]

a greeting to all of you I need your advice on what video card to buy using madVR in 4k ... I was focused on this ZOTAC GeForce 1080 GTX AMP 8GB Extreme Edition .. accept any advice .. thanks

I can vouch for that card as it´s the same I have in my 4K setup. It comes with an excellent cooler (average temp 65C) and high frequencies out of the box (typically 2050-2100 Mhz).

Having said that:

- All GTX 1080 models perform more or less the same, with a 100-150 Mhz max variance depending on the quality of the cooler (Pascal is pretty temp sensitive). If you´re not also gaming, just buy the cheapest version you can find. Games @ 4K will stress the card much further than madvr.

- A 1060, 1070 or any other card will work as well in 4K, they will be just somewhat slower in rendering time. It all depends on how important is to you to crank up the quality settings in madvr. I had a 980ti previously (= 1070) in the same 4K setup and run without problems. I wouldn´t upgrade only for madvr. I would do it if you´re going to game in 4K as well.

- If money is not a factor, wait until 1Q'17 and buy instead a 1080Ti, which will be around 25-30% faster than the 1080. Expected to be a 850-900$ card, though.

[Neo-XP]

Ok so I've jumped onto the latest madvr with this new NGU scaler (which is excellent), but I must have missed something about the change of all the settings. Where did image doubling go? I understand that the settings dialogues have been changed but I can't find any info from madshi in his latest updates in this thread.

My point is, I find the new settings kind of confusing, that's all. Especially with regards to the image doubling.

If you choose to use NGU for luma doubling (and you will use it, because it is excellent), you won't be able to select the algorithms that you desire for chroma upscaling and luma upscaling/downscaling anymore, to use after the luma doubling.

For instance, NGU can't be used to upscale 720p to UHD with my config anymore, because of the rendering times of the forced algorithms to have approximately the same quality. But you can use the v0.91.1 version to pick the algorithms that you want to fine-tune the image quality vs rendering times for your configuration.

I have a perfect result with those with my little 970 GTX :

chroma upscaling : super-xbr - sharpness: 150 (only NNEDI3 and NGU give better results, but rendering times are awful)
image downscaling : Catmull-Rom (if you want soft) or Lanczos 3 (if you want sharp) [both AR + LL], but I prefer sharp
image doubling : double luma resolution - 2x supersampling - NGU, medium quality
image upscaling : Lanczos 3 AR (sharper than Jinc, and very good to use after luma doubling to upscale anything < FHD to UHD)
upscaling refinement : add grain 3-4 (new refinement, still testing which is best, 4 for now)
artifact removal : reduce banding artifacts (high/high) + reduce ringing artifacts only (reduce dark halos remove too many details)
dithering : Ordered Dithering + use colored noise + change dither for every frame (good rendering times and almost identical to Error Diffusion)

[har3inger]

I am also noticing far faster rendering on Dx9 compared to dx11--it's the difference between not being able to run NGU high at all and running NGU high comfortably.

Edit: Disregard that. After switching back and forth a couple times, they're the same (faster) speed now. No idea why.

Is there any reason not to use dx9, given that it seems more efficient? I'm guessing it may be faster on older GPU architectures.

I'm using NGU in conjunction with "reduce ringing artifacts". I definitely notice less source haloing and haven't really found enough ugly interactions between the two.

[AngelGraves13]

I'm also not a fan of bicubic for downscaling from NGU. Jinc is better

[pose]

I'm also not a fan of bicubic for downscaling from NGU. Jinc is better
Jinc was not available for downscaling in previous builds as well.

[StinDaWg]

Is there any reason not to use dx9, given that it seems more efficient?
For one, 10 bit is only available on D3D11 fullscreen exclusive. I don't know if it makes a huge difference, but IIRC it's supposed to help prevent banding.

[Asmodian]

Jinc was not available for downscaling in previous builds as well.

Jinc was added for downscaling a few versions ago (v0.90.12).

[Neo-XP]

Jinc was added for downscaling a few versions ago (v0.90.12).

Exactly.

Here you can see a comparison between Jinc, Lanczos3, Bicubic150 and SSim1D100 for downscaling after 2x supersampling with NGU:

https://s16.postimg.org/jm46cg02p/Jinc_AR.png (https://postimg.org/image/jm46cg02p/) https://s16.postimg.org/ytk5wsrxd/Lanczos_3_AR.png (https://postimg.org/image/ytk5wsrxd/) https://s16.postimg.org/wnpv2aogx/Bicubic150_AR.png (https://postimg.org/image/wnpv2aogx/) https://s16.postimg.org/nwiu8155t/SSim1_D100_AR.png (https://postimg.org/image/nwiu8155t/)

And zoomed:

https://s16.postimg.org/pgqu38035/Jinc_AR.png (https://postimg.org/image/pgqu38035/) https://s16.postimg.org/583ghi0s1/Lanczos_3_AR.png (https://postimg.org/image/583ghi0s1/) https://s16.postimg.org/q2fszbv5d/Bicubic150_AR.png (https://postimg.org/image/q2fszbv5d/) https://s16.postimg.org/4haqbpyep/SSim1_D100_AR.png (https://postimg.org/image/4haqbpyep/)

Make your choice. For me it is Lanczos3 because Bicubic150 has too much aliasing and Jinc is a little bit too soft.

[har3inger]

For one, 10 bit is only available on D3D11 fullscreen exclusive. I don't know if it makes a huge difference, but IIRC it's supposed to help prevent banding.

Right, I forgot about that. Thanks for reminding me.

10 bit output only benefits you if your display/GPU/cable supports 10 bit. Otherwise you are far better using the native bit-depth and good madvr dithering than whatever hardware/GPU dithering/truncation your setup would perform in 10bit mode.

[JarrettH]

I think cubic was chosen because - back 10 pages ago - it was the difference between being able to run ngu-low and it being unusable.

Maybe Catmull should be used in place of bicubic150? I've never a/b them though. I thought we were all happy with Catmull, especially as as downscaler, for a long time :cool:

[pose]

Jinc was added for downscaling a few versions ago (v0.90.12).
I missed that, thanks. Was it worth to use with high quality NNEDI3 or NGU settings over SSIM?

[Uoppi]

I was sceptical at first but I must say I really like the preset approach.

Letting madvr choose what's "good enough" for a given situation - based on user-definable quality settings - is a welcome improvement in my books.

I can actually see myself scrapping about half of my profiles, which were a convoluted mess to set up, not to mention test each profile every time some algo or setting got fine-tuned. I don't have the time anymore (wife would say I NEVER had the time lol).

If madvr wants to expand beyond being a niche product, simplifying the UX is vital and madshi is on the correct path.

[aufkrawall]

Make your choice. For me it is Lanczos3 because Bicubic150 has too much aliasing and Jinc is a little bit too soft.
Thx for the example.
To me eyes, the downscaled 1080p picure doesn't look aliased with bicubic 150 on my 109ppi display. It just looks sharper to me.

[Warner306]

I was sceptical at first but I must say I really like the preset approach.

If madvr wants to expand beyond being a niche product, simplifying the UX is vital and madshi is on the correct path.

Agreed, the performance optimizations possible make presets worthwhile. Suddenly, everyone can run NGU high by choosing the right companion settings.

My own settings are much faster with the given presets.

[madshi]

I've heard a lot of feedback about the new settings dialog organization. Thanks, guys.

But I don't remember anybody commenting yet on the new NGU-Low quality level. It's faster but softer than the old NGU-Low. But is it still "acceptable"? How's sharpness, aliasing, ringing etc for you? Is it worthy of the "NGU" sticker?

[jkauff]

But I don't remember anybody commenting yet on the new NGU-Low quality level. It's faster but softer than the old NGU-Low. But is it still "acceptable"? How's sharpness, aliasing, ringing etc for you? Is it worthy of the "NGU" sticker?
I would guess that those of us with fairly powerful GPUs haven't even tried it. Even doubling a Blu-ray can be done with NGU-Medium on my GTX 1060.

[Anima123]

I've heard a lot of feedback about the new settings dialog organization. Thanks, guys.

But I don't remember anybody commenting yet on the new NGU-Low quality level. It's faster but softer than the old NGU-Low. But is it still "acceptable"? How's sharpness, aliasing, ringing etc for you? Is it worthy of the "NGU" sticker?

I use NGU-Low a lot with my old nVidia 640M, and am happy with image quality it brings with low GPU consumption.

Low in current version can playback 720p -> 1080p with a little bit higher rendering time than 576p -> 1080p with Old Low.

Considering AR of consequent downscaling cannot be disabled anymore, it would have been more faster.

[aufkrawall]

But I don't remember anybody commenting yet on the new NGU-Low quality level. It's faster but softer than the old NGU-Low. But is it still "acceptable"? How's sharpness, aliasing, ringing etc for you? Is it worthy of the "NGU" sticker?
I need to be able to quadruple with it in order to judge. This is not possible with the new profiling.

[thighhighs]

Multi-GPU is a complicated mess and everyone is giving up on it for a reason - game devs and AMD/nvidia are all dropping the ball. Only DX12 could save it but that requires extra work (and training) for the game/app developers. I think only ONE game has dx12 multi GPU support as of right now (the last Deus EX IIRC? and that's not a really well optimized game to begin with).

Just look at how few recent releases support multi-GPU and how bad it is most of the time when they actually do. Poor scaling, bugs, (micro)stuttering etc...

This is even more ridiculous than asking to support VRR for video playback IMO.
Sounds like SLI always bad. Games like BF4 have perfectly multi-gpu scaling (http://www.gamegpu.com/images/stories/Test_GPU/Action/Battlefield_4/test/1920_msaa.jpg) and smooth gameplay.
I don't play all latest games, some games support sli, some not, latest watch dogs 2 support it. And madvr not sli-friendly now, i get better performance when sli disabled, do not see anything ridiculous here, some would like see changes here if it is technically possible.

[Warner306]

Haven't had to use NGU low, other than for chroma upscaling. Hard to tell if it's working well or not.

[ryrynz]

Haven't had to use NGU low, other than for chroma upscaling. Hard to tell if it's working well or not.

Based on my own tests super-xbr AR (100+) is a better choice for chroma over NGU low at least right now quality wise.
NGU low comes off a bit messy for chroma. The 'very high' is quite nice though, but the same issues are there with it as for luma, it brings unwanted details out into the open.
Once it gets it's luma guidance it might just become the best overall chroma upscaler but the performance impact is significant in comparison to other options.

[fedpul]

I've heard a lot of feedback about the new settings dialog organization. Thanks, guys.

But I don't remember anybody commenting yet on the new NGU-Low quality level. It's faster but softer than the old NGU-Low. But is it still "acceptable"? How's sharpness, aliasing, ringing etc for you? Is it worthy of the "NGU" sticker?

madshi, I have set it for my dad's notebook SD profile and I think it is a good option for low power graphics or integrated cards. Quality is really good for the performance hit. (HD 4000 on i7). But I haven't done comparisons.

[fedpul]

Exactly.

Here you can see a comparison between Jinc, Lanczos3, Bicubic150 and SSim1D100 for downscaling after 2x supersampling with NGU:

https://s16.postimg.org/jm46cg02p/Jinc_AR.png (https://postimg.org/image/jm46cg02p/) https://s16.postimg.org/ytk5wsrxd/Lanczos_3_AR.png (https://postimg.org/image/ytk5wsrxd/) https://s16.postimg.org/wnpv2aogx/Bicubic150_AR.png (https://postimg.org/image/wnpv2aogx/) https://s16.postimg.org/nwiu8155t/SSim1_D100_AR.png (https://postimg.org/image/nwiu8155t/)

And zoomed:

https://s16.postimg.org/pgqu38035/Jinc_AR.png (https://postimg.org/image/pgqu38035/) https://s16.postimg.org/583ghi0s1/Lanczos_3_AR.png (https://postimg.org/image/583ghi0s1/) https://s16.postimg.org/q2fszbv5d/Bicubic150_AR.png (https://postimg.org/image/q2fszbv5d/) https://s16.postimg.org/4haqbpyep/SSim1_D100_AR.png (https://postimg.org/image/4haqbpyep/)

Make your choice. For me it is Lanczos3 because Bicubic150 has too much aliasing and Jinc is a little bit too soft.

Bicubic 150 and SSIM 1D are both sharper than Jinc and Lanczos 3. For them all I think I prefer Bicubic 150 even more than SSIM 1D looking at the zoomed pics. But please think that you were not be able to notice such differences between them at normal viewing conditions. Plus Jinc costs a lot ms for downscaling compared to SSIM 1D and even more with Bicubic 150.

[AngelGraves13]

I think the goal for madshi is to make NGU just an option like Jinc without any drop down menus for upscaling. It'll likely all be automatic at some point and the OSD will likely remove any info on the downscaling steps involved. The quality needs to improve a bit more before that can happen. It's just too sharp at the moment and needs a variable level of soften edges depending on the level of NGU.

I actually wouldn't mind it being that way and behaving more as a scaler than a doubler, but we'll just have to wait and see.

[har3inger]

NGU low is a definite improvement over the other "dumb" scalers (jinc, bicubic, lanczos). I think it's the go-to scaler for low-demand laptop playback.

Otherwise, if you can run it, I find NGU high (the new high, old medium) is a breakpoint where the image becomes noticeably better.