madVR - high quality video renderer (GPU assisted)

[RyuzakiL]

Originally Posted by xabregas

I see better detail retention with super-xbr, in the windows, no matter the ringing, sometimes ringing is part of the source. Windows have much more detail

And that's the reason why i stick to S-XBR, I think S-XBR is the most HTPC friendly algo among the rest.

In my opinion when building an HTPC Performance should come first before quality, since running Full NNEDI3 renders gpu to run at its fullest and depending on the gpu cooler, noise problems. And thankfully S-XBR was created to cater for many HTPC's that don't have powerful gpu's. and It seems NNEDI3 consumes alot of GPU load for less obvious improvement in picture quality over S-XBR which oozes with efficiency and better detail retention.

Unless one enjoys looking at static picture just to notice a minute improvement of NNEDI3 then by all means. enjoy staring.

[ryrynz]

Originally Posted by xabregas

I see better detail retention with super-xbr, in the windows, no matter the ringing, sometimes ringing is part of the source. Windows have much more detail

Prefer NNEDI3 in that example, more natural.

[aufkrawall]

Originally Posted by RyuzakiL

Unless one enjoys looking at static picture just to notice a minute improvement of NNEDI3 then by all means. enjoy staring.

Unnecessary trolling, the differences can be seen in motion instantly.

[leeperry]

After quite a bit of fiddling today, I think there's no free lunch to be expected. Many sources come with ringing to begin with and yeah, if you magnify it it will become more obvious...sxbr digs more than NNEDI3 IMHO and as a result also rings more, I currently run sxbr50 with SR(3 passes, 0.65 sharpness, 0.01 softness) and it's working really well IME in mVR .15, OTOH 0.00 softness looks too edgy to my taste.

[baii]

Originally Posted by aufkrawall

Suggestion: Could we have a checkbox for image enhancement to just enable it if no luma upscaling is applied?
This would be useful to prevent sharpening of sharpening artifacts when also using it as an upscaling refinement.
I think it would be neat to help NNEDI3 with AdaptiveSharpen, but I don't want it combined AdaptiveSharpen of image enhancements, which on the other hand is neat to fight chroma blur a bit.

A workaround is to make profile with source and target dimension as scalefactor seem to be a little wonky. Since the profile can't use parameter compare to parameter now(say source width<target width), it can be a little cumbersome, but you can make it work.

Sent from my 306SH

[madshi]

I'll reply to all your posts later. For now I would just like to share a new test build, using a tweaked SuperRes algorithm:

http://madshi.net/madVR8816b.rar

SuperRes image quality should generally be slightly better than in the official v0.88.16 build, furthermore I've added an anti-ringing filter which should noticeably reduce ringing artifacts introduced by SuperRes. I hope that the anti-ringing filter will allow us to use algo 2, which is my favorite algorithm, but didn't work well due to the added ringing artifacts in v0.88.16. FWIW, if I had to decide on final SuperRes configuration right now, it would be this:

- algo = 2
- use alternative color space = off
- low: strength=0.5; passes=1
- medium: strength=1.0; passes=1
- high: strength=1.0; passes=2
- ultra: strength=1.0; passes=4

Would like to have your feedback about this. If you believe that the above settings are a totally bad idea, please let me know. Ideally with a small sample which shows that other settings work much better. Or if you like my settings, please let me know, too. Thanks!

[Warner306]

Originally Posted by JarrettH

When image doubling, does chroma upscaling happen before chroma doubling?

In that case, would the chain look like...

chroma > super-xbr (chroma upscaling) > super-xbr (chroma doubling)

Yes, it works just like that.

[AngelGraves13]

What kind of ringing or aliasing would an anisotropic filter produce? Assuming it could be used for upscaling chroma...

[e-t172]

Hey, quick question. I am planning on building a rig based around a UHD screen and a 980Ti (not just for movies). Of course, since there is little UHD content in the wild I intend on primarily watching upscaled 1080p movies. Up until now I didn't care about upscaling too much with my current config, but of course in this project upscaling quality will be paramount, since my viewing angle will be optimized for UHD (i.e. the picture will be much bigger), not 1080p. Is NNEDI3 1080p -> 2160p feasible with a high-firepower GPU such as the 980Ti or is it too extreme still? Is NNEDI3 good enough to provide high quality presentation of 1080p upscaled to 4 times the typical effective picture size?

[Anima123]

Originally Posted by madshi

it would be this:

- algo = 2
- use alternative color space = off
- low: strength=0.5; passes=1
- medium: strength=1.0; passes=1
- high: strength=1.0; passes=2
- ultra: strength=1.0; passes=4

Would like to have your feedback about this. If you believe that the above settings are a totally bad idea, please let me know. Ideally with a small sample which shows that other settings work much better. Or if you like my settings, please let me know, too. Thanks!

As I said, madshi, would you please keep at least passes and strength options, and use 3-4 button for default values? In that way, everyone would be happy, without compromise your intention of make madVR easy to use that much.

Edit: Right now, my optimal settings for 1024x576 -> 1920x1080 of SuperRes is 2-passes with strength 0.40, algo 1.

[MS-DOS]

Originally Posted by madshi

I'll reply to all your posts later. For now I would just like to share a new test build, using a tweaked SuperRes algorithm:

http://madshi.net/madVR8816b.rar

SuperRes image quality should generally be slightly better than in the official v0.88.16 build, furthermore I've added an anti-ringing filter which should noticeably reduce ringing artifacts introduced by SuperRes. I hope that the anti-ringing filter will allow us to use algo 2, which is my favorite algorithm, but didn't work well due to the added ringing artifacts in v0.88.16.

What can I say? It's amazing. Even 10 passes don't ruin the image to my personal taste. It's definitely not what I'd use, though.

Originally Posted by madshi

FWIW, if I had to decide on final SuperRes configuration right now, it would be this:

- algo = 2
- use alternative color space = off
- low: strength=0.5; passes=1
- medium: strength=1.0; passes=1
- high: strength=1.0; passes=2
- ultra: strength=1.0; passes=4

Would like to have your feedback about this. If you believe that the above settings are a totally bad idea, please let me know. Ideally with a small sample which shows that other settings work much better. Or if you like my settings, please let me know, too. Thanks!

Either these or just a "passes" option. "High" settings is what I've been using since .16
I always disliked sharpening in general, but this new SuperRes does wonders combined with S-XBR

[AngelGraves13]

Originally Posted by e-t172

Hey, quick question. I am planning on building a rig based around a UHD screen and a 980Ti (not just for movies). Of course, since there is little UHD content in the wild I intend on primarily watching upscaled 1080p movies. Up until now I didn't care about upscaling too much with my current config, but of course in this project upscaling quality will be paramount, since my viewing angle will be optimized for UHD (i.e. the picture will be much bigger), not 1080p. Is NNEDI3 1080p -> 2160p feasible with a high-firepower GPU such as the 980Ti or is it too extreme still? Is NNEDI3 good enough to provide high quality presentation of 1080p upscaled to 4 times the typical effective picture size?

Wouldn't recommend it. It's very taxing and much too sharp. I even find Jinc for upscaling to be too sharp as it makes 720p look overprocessed when upscaled to 1440p. I'd go with Jinc for Chroma and Bilinear for image upscaling. That's just my opinion though.

[madshi]

Originally Posted by Anima123

Edit: Right now, my optimal settings for 1024x576 -> 1920x1080 of SuperRes is 2-passes with strength 0.40, algo 1.

In my tests 10 passes with strength 0.10 looks almost identical to 1 pass with strength 1.00. So 2 passes with strength 0.40 should look almost identical to 1 pass with strength 0.80. If you disagree, please show me a screenshot where 2 passes with 0.40 look clearly better than 1 pass with 0.80.

This is extremely important, because different strength values don't cost performance. But 2 passes is exactly twice as slow as 1 pass.

[AngelGraves13]

I can't find a good reason to use SuperRes. It's just too sharp. I'd rather have a bilinear filtering pass on the image to soften it up to hide source artifacts.

[Anima123]

I just tried between 2 passes, strength 1.0 vs. 4 passes, strength 0.50 with madVR latest test build, result as follow:

https://www.dropbox.com/s/heo0bzw64t...h1.00.png?dl=0

https://www.dropbox.com/s/x94w0few0d...h0.50.png?dl=0

For a moment I almost doubt that I am looking at the same picture. I would say, that madshi has changed the SuperRes's behavior quite a lot from the original.

The original idea is that the images should converge as passes increasing.

Do you have any comment on that madshi?

[aufkrawall]

Neat, new SuperRes has an AA effect, even though the image gets sharper (e.g. with that "high" preset).

But very interesting: The effect of SuperRes is totally different between super-xbr and NNEDI3: It seems to be much stronger with the latter one. Is this by design?

I still think I don't need SuperRes anymore now that AS can be used with NNEDI3.
At least I don't have performance for both, and I really rather enjoy NNEDI3 (you probably guessed it )

super-xbr 75 quadrupling + "high" SuperRes:


NNEDI3 64 + super-xbr 75 quadrupling + AS UR 0.2:

[xabregas]

Originally Posted by aufkrawall

Unnecessary trolling, the differences can be seen in motion instantly.

I agree, i can easily see the differences betwenn algo in motion. For me NNEDI3 seemed the better option for chroma until super-xbr come. Speaking of gpu intensive, super-xbr is better. The PQ with NNEDI3 only gets much better with 64 neurons which is too intensive for me ~38ms rendering. I get ~4ms rendering with super-xbr ffs. I see no amazing differences from NNEDI3 32 neurons and super-xbr 100 for 1080p video on 1080p screens

[aufkrawall]

Originally Posted by xabregas

I see no amazing differences from NNEDI3 32 neurons and super-xbr 100 for 1080p video on 1080p screens

I even can't distinguish between 1080p -> WQHD NNEDI3 64 and 1080p -> WQHD Jinc3 AR (at least with the samples I watched). I think the default settings to only double resolution with scaling factors >= 1,5x totally make sense.

[madshi]

Originally Posted by reverepink

Recommend me some settings for video playback, please.

Please use one of the many madVR guides/tutorials out there. Your question is too general to give a good answer to, and this kind of question gets asked all the time.

Originally Posted by Eyldebrandt

Honestly, the differences between the algorithms and the color space on HQ content @ 4K or WQHD are insignificants.
And with low strength, the action is invisible.

I'm confused, because I think SuperRes is better than the 1st version, but settings are gone.

On the very 1st build,I used superRes with high quality instead of mid, and with everything @ 0.0 except strength.
Since, I never succeed to reach the same "quality" i had at the beginning.

I suppose with "at the beginning" you mean the algorithm which had all the options like anti-aliasing, anti-ringing etc? That version of the algorithm was technically inferior to the latest algorithm. I understand you liked it, but Shiandow has trashed it for the new algorithm. Maybe the new algo can still be improved a bit, but I agree with Shiandow that it's better than the old one.

Originally Posted by leeperry

Oh, do I seriously have to justify myself on this?

No, you don't have to. All I was trying to say is that judging image quality can be a difficult task. And I was trying to prove it by showing that you yourself are changing your opinion once in a while. Anyway...

Originally Posted by leeperry

So HQ looks better or worse depending on the footage and it's now forced, bleh...you even said yourself that you were not sure whether it decreased or improved PQ when you first implemented it.

Yes, but since then I've done some tests and found that I always clearly preferred HQ on. And it's technically, scientifically better, too. And the majority of users seemed to agree, too.

HQ off produces a different look, one that I personally find very artificial looking.

Originally Posted by leeperry

Well, SR only takes care of upscales AFAIK and pretty much all 720p movies are reencodes in one way or another. I'd even dare saying that 99% of them are 1080p downscales. I do see the very same "hard" edges on whatever content when using NNEDI3 or sxbr though, it's fantastic for tiny videos but they both seriously try too hard for 720p@1080p to my eyes. NEDI is sharper than J3AR and yet doesn't give that nasty computerized look, please don't toss it

I've not planned to toss NEDI yet. I know that there are some users who like it.

Originally Posted by Nevilne

quick hard sample for super-xbr chroma ar and bilateral chroma:

start at 150 sharpness, toggle between super-xbr fast ar and slow ar.

madvr.avs

BlankClip()
Subtitle("ONE, TWO, THREE, OUR CHROMA", x=-1, y=150, size=36, spc=4, font="Microsoft Sans Serif", text_color=$aa0000, halo_color=$262626)
converttoyv12.sharpen(0.4)

you can see that fast chroma ar is superior on this image, perhaps it could be a quality/performance option for super-xbr luma/image.

Interesting method for testing!

The slow AR algo in this image believes that allowing ringing is the way to go. Which seems to not be a good decision in this case. But in many other cases the fast AR looks worse than the slow AR. You have the choice now.

For luma/image doubling I believe the slow AR is a clear winner. Tested on several images and slow AR always looked better.

Originally Posted by Anima123

I just tested the new version with NEDI as the image doubler, passes 10, strength 0.30, and use alternative color space checked.

Please try 3 passes with strength 1.00. Does 10 passes with strength 0.30 look better to you? If so, please show some comparison screenshots.

Originally Posted by Anima123

algo 0 show some blocky effects on image and the quality is quite low. The higher the number, more ringing can be observed around the hard-coded subtitles in the image. I would choose algo 1 if there should be only one left.

It seems that algo 1 (algo 0 not counted) is the most ringing resistant one in case more passes are applied. I am happy with the result of 10 passes and 0.30 (maybe someone like less) with algo 1.

Does any of this change with the new test build? The ringing should be less now, so maybe you can use algos with a higher number now?

Originally Posted by SithUK

Madshi: you recommended I try bilinear chroma upscaler to try and resolve my issues with using a 3d lut on an old laptop. I mentioned I watch 1080p content on a 1080p screen. Is there a performance impact from chroma upscaling when watching 1080p at 1080p, ie if there is no upscaling?

I usually don't ask users to test things that make no sense.

Originally Posted by SithUK

What is the best way to read performance using the ctrl+j overlay. Render time? Or is it to use cpuz to monitor gpu and cpu load?

Render times are a good indicator, but don't take them as gospel. E.g. if the GPU clocks up/down, render times are useless.

Originally Posted by leeperry

0.00 sharpness is less forgiving, I kinda liked slightly increasing it in order to hide compression artifacts but yeah OK unforgiving is good too and I'm currently sitting 80cm away from a 3500:1 32"....I guess 0.00 would be just fine from a 3 meters distance.

You mean softness, not sharpness, right?

If there are compression artifacts, that should be handled by a different filter. Which sadly doesn't exist yet.

Originally Posted by leeperry

I still far prefer HQ off in .15, enabling it utterly veils the picture to me. Major bottleneck at work, this is a definite no-go

Try adding some AdaptiveSharpen, that makes the image look more like HQ off.

Originally Posted by Schwartz

I've never seen the appeal of using an upscaler and then slapping on a 'corrective' AR algorithm. Ideally, a scaler shouldn't introduce ringing in the first place. Maybe it's so popular because the red and green bars suggest that it's the best. Remember that these bars aren't an objective assessment. I'd be curious to see each scaler's wave diagram instead. If you want to try something new, there's plenty of other upscalers. I'm a big fan of Spline 3 for upscaling and Mitchell for downscaling. SoftCubic 70-80 is great for dealing with bad encodes and old SD content.

Ideally an upscaler would work magic. In real life we have to make do with what science has given us so far. If you want an algo which doesn't ring by itself, try Bilinear, Nearest Neighbor or Gaussian. Ouch. Pretty much everything else rings. Even NNEDI3 adds a bit of ringing in some situations. As Hyllian already mentioned, some ringing is sometimes beneficial, which makes the whole thing difficult. E.g. try some frequency burst test patterns without ringing. You'll get very bad results.

Originally Posted by Braum

Madshi do you plan on introducing a frame interpolation function in Madvr ?

No. At least not any time soon. There are devs which work on that, see SVP. I doubt I could do a better job than them, at least not with the limited amount of time I have.

Originally Posted by aufkrawall

Suggestion: Could we have a checkbox for image enhancement to just enable it if no luma upscaling is applied?
This would be useful to prevent sharpening of sharpening artifacts when also using it as an upscaling refinement.
I think it would be neat to help NNEDI3 with AdaptiveSharpen, but I don't want it combined AdaptiveSharpen of image enhancements, which on the other hand is neat to fight chroma blur a bit.

My thinking was that image enhancements should only be activated on demand if you happen to play a source which is overly soft. As such, it would still make sense to run image enhancements and upscaling refinement at the same time.

But all of this is still up for discussion. We're not there yet, though. I want to first reduce all the sharpening options as much as possible. Only afterwards we'll go back to see when to use image enhancements vs upscaling refinement, in combination or not etc...

Originally Posted by aufkrawall

Some more NNEDI3 vs super-xbr doubling test. Sorry about the image size, I don't know how to crop always exactly the same pixels.
However, I could do this in future if someone explained this to me.
720p -> WQHD, Jinc3AR chroma

super-xbr 100 is sharper than NNEDI3 64, but ringing gets more obvious and lines aren't as clean.

Yes, NNEDI3 is still king. Totally agree there. But it's also still very slow.

Originally Posted by aufkrawall

I just tried a bit around with AdaptiveSharpen vs. FineSharp as upscaling refinements, and I have to say that I now understand why some people describe FineSharp as pure destruction.
It increases ringing in a very unpleasant way, but you need need a lot of strength to make it actually sharpen things noticeably.
With AdaptiveSharpen, you can use much smaller values and it will really sharpen areas which actually need the sharpen. It also doesn't suffer by ringing like FS in LL.

FineSharp by design sharpens the heck out of image details, which means it also sharpens the heck out of noise and image artifacts. I guess FineSharp should only be used on very clean high quality sources.

Originally Posted by AngelGraves13

I'm really liking super-xbr at 50 for chroma and image doubling with bilinear upscaling/downscaling. Upscaling DVDs with Jinc AR looks pretty terrible, but Bilinear hides the source artifacts quite well. Sharpness is nice, but not if it makes the videos look worse.

Bilinear? Euwhhh! If you want to hide source artifacts, I'd suggest SoftCubic instead.

Originally Posted by aufkrawall

At least I wouldn't choose algo 2 because of the softness and between 1 and 3, I don't see a real difference with 720p content I've quickly thrown at it. But since 1 tends to make some colors to bright with cartoons, maybe there's no good reason not to use 3?

Softness? Algo 2 should make lines thinner compared to algo 0 or 1. On a quick check this may look softer. But it also looks more natural, IMHO. Might be a matter of taste, though. If you like fat lines, probably AdaptiveSharpen is also your friend.

Originally Posted by JarrettH

When image doubling, does chroma upscaling happen before chroma doubling?

In that case, would the chain look like...

chroma > super-xbr (chroma upscaling) > super-xbr (chroma doubling)

madVR always upsamples chroma to the same resolution as luma first (4:2:0 -> 4:4:4).

Originally Posted by AngelGraves13

What kind of ringing or aliasing would an anisotropic filter produce? Assuming it could be used for upscaling chroma...

I think anisotropic filtering is for 3D stuff, like surfaces that are perspectively distorted. I don't think it's useful for video processing.

Originally Posted by e-t172

Hey, quick question. I am planning on building a rig based around a UHD screen and a 980Ti (not just for movies). Of course, since there is little UHD content in the wild I intend on primarily watching upscaled 1080p movies. Up until now I didn't care about upscaling too much with my current config, but of course in this project upscaling quality will be paramount, since my viewing angle will be optimized for UHD (i.e. the picture will be much bigger), not 1080p. Is NNEDI3 1080p -> 2160p feasible with a high-firepower GPU such as the 980Ti or is it too extreme still? Is NNEDI3 good enough to provide high quality presentation of 1080p upscaled to 4 times the typical effective picture size?

A lot depends on the frame rate, of course. 60p processing consumes 2.5x as much performance as 24p. Doing NNEDI3 1080p -> 2160p should be easy for a 980Ti with 24p. Not sure about 60p.

But yes, NNEDI3 is a good choice if you want max quality and it's the best you can use for that job. I'd add one pass of SuperRes or FineSharp on top of that, to improve perceived sharpness even more. You may also want to use debanding because many Blu-Rays still have banding in them.

If you want to save GPU power, super-xbr would be a *MUCH* faster alternative. Quality is a bit lower than NNEDI3, but still quite good.

Originally Posted by MS-DOS

What can I say? It's amazing. Even 10 passes don't ruin the image to my personal taste. It's definitely not what I'd use, though.

Any obvious changes compared to the previous build?

Originally Posted by Anima123

I just tried between 2 passes, strength 1.0 vs. 4 passes, strength 0.50 with madVR latest test build, result as follow:

For a moment I almost doubt that I am looking at the same picture. I would say, that madshi has changed the SuperRes's behavior quite a lot from the original.

The original idea is that the images should converge as passes increasing.

Do you have any comment on that madshi?

So you agree that 2 passes 1.0 looks very similar to 4 passes 0.5? So the settings I suggested do make sense, right?

Originally Posted by aufkrawall

Neat, new SuperRes has an AA effect, even though the image gets sharper (e.g. with that "high" preset).

But very interesting: The effect of SuperRes is totally different between super-xbr and NNEDI3: It seems to be much stronger with the latter one. Is this by design?

I don't know. So you'd be fine with my suggested SuperRes settings?

[Anime Viewer]

Originally Posted by Anima123

I just tried between 2 passes, strength 1.0 vs. 4 passes, strength 0.50 with madVR latest test build, result as follow:

https://www.dropbox.com/s/heo0bzw64t...h1.00.png?dl=0

https://www.dropbox.com/s/x94w0few0d...h0.50.png?dl=0

For a moment I almost doubt that I am looking at the same picture.

The pictures you posted look almost identical to me. Aside from the creases in his forehead and the creases in the sleeve of the passerby I'm not seeing anything noticeably different between the two. Of the two I think the first one (strength 1 with two passes) looks better than the second one (strength 0.50 with four passes).

Originally Posted by madshi

In my tests 10 passes with strength 0.10 looks almost identical to 1 pass with strength 1.00. So 2 passes with strength 0.40 should look almost identical to 1 pass with strength 0.80. If you disagree, please show me a screenshot where 2 passes with 0.40 look clearly better than 1 pass with 0.80.

This is extremely important, because different strength values don't cost performance. But 2 passes is exactly twice as slow as 1 pass.

Given both the performance difference and the small difference I see between the two pictures Anima posted I vote for the higher strength with less passes. In my testing I see very little ( on my sources) in the low, medium, high, and ultra settings you noted before. I could see myself using the low or medium settings, but don't think I'd every use the high or ultra settings given how much of a performance hit they are with not enough significant image improvement. (Pretty much the same reason why I choose to use Super-xBR instead of NNEDI3).

Edit: I like the way it combines with image enhancements (not the upscaling refinement version) of Adaptive Sharpening (not sure on the strength...0.3 or 0.5 maybe), but I'll hold back on further commenting on that until you want to talk about combining effects.