madVR - high quality video renderer (GPU assisted)

[pirlouy]

FWIW I'm ok with the new interface config. The "automatic" option is a really nice idea for casual users like me.

[huhn]

Originally Posted by JarrettH

If you choose downscale quality: low, madshi has left you with bilinear, which should hide source artifacts. I think he chose this intentionally to satisfy the SD people.

i don't think that helps...

and SD people? i choice the stuff i watch based on the content not on resolution...

[leo91531]

test@GTX1080 (RX480 not yet)


This version feels more simple to adjust, but AMD's graphics card seems to have some problems.

madvr@RX400 bug?
https://youtu.be/QlVLuGVmZjo
AMD GPU load is unstable,GPU load more than 90%(not Avg load) then video easy to drop frame.(not power saving mode)

RX480 GPU load


GTX1080 GPU load

[aufkrawall]

Originally Posted by leo91531

madvr@RX400 bug?
https://youtu.be/QlVLuGVmZjo
AMD GPU load is unstable,GPU load more than 90%(not Avg load) then video easy to drop frame.(not power saving mode)

Did you try software decoding? Had a similar issue with a R9 390 and 1080p 60fps H.264.

[Warner306]

I vote to remove downscaling quality or at least add an option to use the downscaling algorithm.

It was actually easier to understand before. Chroma quality is the only box that should be available in a subgroup called "image upscaling." I don't think downscaling belongs. This adds complexity.

Perhaps the image downscaling setting can appear as automatically selected when hovering over NGU or NNEDI3? This avoids a situation where the same setting is repeated.

I am of course assuming image downscaling is part of a profile dedicated to image doubling.

[ikarad]

Originally Posted by huhn

nearly every g-sync monitor is HFR so just use smoothmotion.
smoothmotion at 120 hz or more is a totally save.
the audio clock mismatch is fixed this way too.

At 120 hz, smoothmotion is useless, not?

[Neo-XP]

Originally Posted by Warner306

I vote to remove downscaling quality or at least add an option to use the downscaling algorithm.

It was actually easier to understand before. Chroma quality is the only box that should be available in a subgroup called "image upscaling." I don't think downscaling belongs. This adds complexity.

Perhaps the image downscaling setting can appear as automatically selected when hovering over NGU or NNEDI3? This avoids a situation where the same setting is repeated across two downscaling menus.

Agreed. I just want to be able to use Lanczos3 AR again for luma/chroma upscaling and luma downscaling, after NGU. The new settings don't give me the same result as before. This is a regression for advanced users.

The low/medium/high/very high quality settings should be chosen as presets in the algorithms menus, not in the NGU part. NGU should be an option in those menus (a checkbox), not used as a replacement, but as a complement.

The "chroma upscaling", "image downscaling" and "image upscaling" menus are totally useless now when using NGU for luma upscaling (because of the forced presets). I don't understand this decision.

[huhn]

Originally Posted by ikarad

At 120 hz, smoothmotion is useless, not?

no not at all.

most content is 23p (24000/1001 hz) and 25/50 content will judder at 120 hz.
smoothmotion is very useful for 120 HZ screen and like said before audio clock issue are fixed too.

[fedpul]

IMHO, I think what madshi is trying to do is let the options that really change IQ and remove the ones that are super power hungry for almost no visible difference, and I am not talking the difference you can see if you see pixel per pixel at 1cm from the screen, but at a normal view distance. It is like an audiophile that can claim that can hear differences between redbook standard and ultra HD tracks (means no offence, I am just being objective). Probably if madshi shows us two different settings telling which algos are in the presets he shows, and asking us to identify, we could fail or have 50/50 chances to be right. Like blind tests for audio files.

I repeat, this is just my opinion.

[leo91531]

Originally Posted by aufkrawall

Did you try software decoding? Had a similar issue with a R9 390 and 1080p 60fps H.264.

my LAV setting is HW Decoding.


when I ALT+TAB to desktop the rendering time will be down

sorry that GPU-Z Screen is too bright cause recording is not very good. You can look carefully that GPU load curve,it was not smooth and not stable.

[ikarad]

Originally Posted by huhn

no not at all.

most content is 23p (24000/1001 hz) and 25/50 content will judder at 120 hz.
smoothmotion is very useful for 120 HZ screen and like said before audio clock issue are fixed too.

23.976p and not 23p.
I have 120hz and without smoothlmotion there is no judder.
I have tried smoothmotion and reclock to try to see any difference and I have never seen any difference.
In 60hz mode There is judder.


Madshi had said when I have posted this question many months ago that smoothmotoion was useless with 120hz monitor .

Maybe with 25/50 content but I don't have 25/50 content. I speak only about 23.976p content.

[huhn]

even windows and GPU driver are using 23p/23 hz for 24000/1001
and not to forget madVR...

but what ever.

playing a 23p file at 120 hz without reclock or smoothmotion will result in a frame repeate from time to time about all 42 sec. if you don't see that fine for you but it still happens.

something like that: https://abload.de/img/23pproblem.yyuz0.png

[Georgel]

Originally Posted by aufkrawall

Did you try software decoding? Had a similar issue with a R9 390 and 1080p 60fps H.264.

I have the same problem on GTX 1080.

There are some spikes to 200ms on decoding sometimes, but GPU usage is way under 100%, but it doesn't impact frame rate...



GTX 1080 can do everything at max, but on img upscaling luma / chroma / downscale at high on NGU with this version - but cooling is too noisy so it's not a profile to run for everyday enjoyment.

Settings for enjoyment ///

chroma / NG / high
downscling / JINC
UPSC / NGU / luma high, chroma / auto, downscale / medium - activate only if needed

refinement thin edges at 4.2 after every 2X step

I guess that my settings are not half as crazy as I made it seem throughout the time.

There is still a considerable amount of grinding / coil whine that is not there without NGU, but the image quality is way too good to skip.

Now it's the time I watch my poorly encoded Anime Music Videos and old Anime with a better image!

/This thing makes Fate anime series look like image quality is on steroids!

[JarrettH]

Originally Posted by fedpul

IMHO, I think what madshi is trying to do is let the options that really change IQ and remove the ones that are super power hungry for almost no visible difference, and I am not talking the difference you can see if you see pixel per pixel at 1cm from the screen, but at a normal view distance. It is like an audiophile that can claim that can hear differences between redbook standard and ultra HD tracks (means no offence, I am just being objective). Probably if madshi shows us two different settings telling which algos are in the presets he shows, and asking us to identify, we could fail or have 50/50 chances to be right. Like blind tests for audio files.

I repeat, this is just my opinion.

Yep. I mean when you get down to it, for instance with chroma upscaling, super-xbr brings it to a good quality level, but I hope you don't blink (and your nose must be touching the screen) when looking for improvements after

[aufkrawall]

Originally Posted by Georgel

I have the same problem on GTX 1080.

There are some spikes to 200ms on decoding sometimes, but GPU usage is way under 100%, but it doesn't impact frame rate...

You mostly have to stay below 90% GPU usage with madVR to avoid drops. I don't have any spikes with GTX 1070, a lot of load and DXVA2-cb.

[Telion]

Originally Posted by MS-DOS

madshi, anti-ringing checkbox in image downscaling is greyed out for me. Switching to chroma upscaling and back makes it active until I close the settings window.

The same goes for "image upscaling" when non-doublers are selected.

Originally Posted by omarank

I would want to choose the upscaling algorithm to be used after image doubling for further upscaling to the target resolution. That algorithm may also be used for upscaling to lower scaling factors which don’t activate the selected image doubling algorithm. For instance, if NNEDI3 or NGU is selected in upscaling settings to be used, say, when scaling factor is 1.5x or bigger, Bicubic60 AR is used for upscaling to less than 1.5x scale factor whereas I would want to use Jinc there.

+1. Doublers shouldn't be mutually exclusive with non-doublers. I also vote for some kind of a custom/advanced checkbox for doublers which would switch their drop-down content from quality profiles to specific algorithms so their combination can be selected more freely.

[ikarad]

Originally Posted by huhn

even windows and GPU driver are using 23p/23 hz for 24000/1001
and not to forget madVR...

but what ever.

playing a 23p file at 120 hz without reclock or smoothmotion will result in a frame repeate from time to time about all 42 sec. if you don't see that fine for you but it still happens.

something like that: https://abload.de/img/23pproblem.yyuz0.png

With reclock, I have frame repeated about all nearly 5 min.
I can have the same thing or even better (6.3 min for frame repeated) if I create custom resolution with 119.88hz (exact multiple of 23.976).

With 119.88 or 120hz or reclock, I have never seen the difference. Have you a recording to see the difference?

[Crimson Wolf]

Originally Posted by omarank

I too have been using Catmull-Rom AR LL for downscaling after 720p to 1080p doubling, with previous madVR builds. Bicubic150 or SSIM1D look too sharp to my eyes in some cases.


I agree.

I would want to choose the upscaling algorithm to be used after image doubling for further upscaling to the target resolution. That algorithm may also be used for upscaling to lower scaling factors which don’t activate the selected image doubling algorithm. For instance, if NNEDI3 or NGU is selected in upscaling settings to be used, say, when scaling factor is 1.5x or bigger, Bicubic60 AR is used for upscaling to less than 1.5x scale factor whereas I would want to use Jinc there.

Correct me if I'm wrong, but can't you do that with profile groups now? Even with the new GUI? Just set srcWidth <= ??? "bicubic60ar + double scaling with NGU or NNEDI3" and srcWidth > "jinc" (since jinc disables doubling)?

I'm quite liking the new GUI so far.

[kalston]

Originally Posted by Crimson Wolf

Correct me if I'm wrong, but can't you do that with profile groups now? Even with the new GUI? Just set srcWidth <= ??? "bicubic60ar + double scaling with NGU or NNEDI3" and srcWidth > "jinc" (since jinc disables doubling)?

I'm quite liking the new GUI so far.

I think it's not too much to ask to be able to chose which upscaler to use after the doubling like we could before. I'm sure you can find a workaround in the meantime but wasn't this GUI update supposed to streamline things?

Or maybe he should just make JincAR the default instead of bicubic60, there are still some integrated GPUs that can choke on it (my Surface 3's with some content) but it should be fine for the vast majority of users.

[Telion]

madshi, since v0.91.2 I have a problem when NGU image upscaler uses Bicubic60 to process chroma. The picture is a mess of coloured rectangles and mostly freezed while I can see that the luma part is moving. Switching between 100% window and windowed FS modes or changing the size of a window, I noticed that the freezed image is based on the last one that was shown correctly in a 100% or below-doubling upscaled window. Even if I restart the player and go FS before opening the same video, the freezed image would be linked to a previous run of the player. If it is a new video, the freezed image would be based on its first frame. If I switch to NGU high and select NGU to also double the chroma there is no such problem, and when using non-doublers either. I can't test NNEDI3 on my ancient rig and it is on XP, so I suppose the problem is related to its old presentation path since no one has reported it yet.

Debug log & screenshot