madVR - high quality video renderer (GPU assisted)

[madshi]

Originally Posted by Xaurus

There is a bug still in 91.7 where you have a random profile, it displays a high or higher rendering time than "frame" but in reality it is lower.

Can't reproduce it. Are you sure it's really a new bug in v0.91.7?

Originally Posted by Arm3nian

Because when you double, sharpen edges is applied, and when you double once more, sharpen edges is again applied. But when you quadruple in one step, sharpen edges is applied only once. I get this from the fact that doubling twice looks similar to quadrupling directly, in my sample. If anything, quadrupling directly looks better. But with sharpen edges and refine after every step turned on, doubling twice clearly looks sharper.

That's all correct, but what does that have to do with 8.0 or 16.0 which you mentioned?

Originally Posted by pirlouy

Too bad I still have to choose "chroma upscaling" and "image downscaling" settings.

For now.

Originally Posted by cork_OS

New UI is very nice, but one thing confuses me: seems NGU AA and NGU Sharp share all upscaling settings. Is it bug or feature?

I wasn't sure if it's better this way or separating the settings. Finally I decided to implement it the way it is now, so you can later quickly switch between e.g. NGU Soft, Standard and Sharp, with all other settings identical, so you can more easily compare which of the 3 different NGU sharpness levels you prefer.

Originally Posted by cork_OS

Is mixed NGU upscaling (AA->Sharp) abandoned or still planned?

I don't know yet. Generally I prefer to work on direct quadrupling algos, because they have a much better quality vs speed ratio.

Originally Posted by har3inger

Yep, the new AR relaxed for downscaling after NGU doubling is what's causing the increased render time. When I set downscale to bicubic 60, all the builds render at the same speed.

This is kind of unfortunate, as the only one not using AR is bicubic 60 (bilinear doesn't count). For me, linear light on downscale after doubling has negligible impact on performance, but relaxed vs strict AR is quite substantial.

Which NGU quality level are you using? Which GPU do you have? Generally, the AR speed difference should not be that dramatic for GPUs with decent speed.

Anyway, you can always use the "use downscaling settings" option, then you have full control.

Originally Posted by Neo-XP

It is a shame that you can't choose not to use strict AR anymore, because the "same as downscaling algo" option is gone.

That's not true. It's available in v0.91.7.

Originally Posted by Georgel

Gotta wonder at this moment though, leaving them inside madVR does affect the complexity of programming?

Does it have any negative impact in your workflow when designing it?

Depends on the feature/option.

[Neet009]

Originally Posted by madshi

madVR v0.91.6 released

Now I'd like to have your feedback about the following topics:

1) Is the new upscaling settings page fine now?
2) When disregarding speed, how would you rate NGU Anti-Alias Very High vs. NNEDI3 256-taps, for image upscaling?
3) When taking speed into account, how would you rate NGU Anti-Alias vs NNEDI3 at the various quality levels, for image upscaling?
4) Same as 2) and 3), but for chroma upscaling.

1. I agree that it maybe too many optoins, but it is fine and someone needs those options. IMO, it is easy to use in general.

2. According to my test, NGU AA very high beats NNEDI3 256 as a whole, especially for low-res aliased anime content. NGU AA makes line more focus and less aliasing as expected.

3. I found that NGU AA very high is slightly faster than NNEDI3 64 when zooming 200% without doing any upsaling or downscaling after luma doubling. However, when doing upsaling or downscaling after luma doubling, NGU AA very high is slightly slower than NNEDI3 64. I don't know why but it doesn't matter. I'll replace NNEDI3 with NGU AA on my settings.

4. For chroma upscaling, it's hard to tell the quality differents between NNEDI3 and NGU AA, but I can confirm that NGU AA med is faster than NNEDI3 16.

Graphics card:NV GTX1060 6GB (driver 373.06)
OS:win7 x64

Anyway, I like NGU AA and NGU sharp(2x and direct 4x), they keep good balance between quality and performance. Again, Thank you for your great works.

[Xaurus]

Originally Posted by madshi

Can't reproduce it. Are you sure it's really a new bug in v0.91.7?

No, it's not a new bug, I said it is still present in 91.7. Apologies if it was unclear.
I can easily reproduce it, question is, how can I assist you with data?

[Bloodred]

Originally Posted by madshi

Now I'd like to have your feedback about the following topics:

1) Is the new upscaling settings page fine now?
2) When disregarding speed, how would you rate NGU Anti-Alias Very High vs. NNEDI3 256-taps, for image upscaling?
3) When taking speed into account, how would you rate NGU Anti-Alias vs NNEDI3 at the various quality levels, for image upscaling?
4) Same as 2) and 3), but for chroma upscaling.

1) Yup, I like the increased control, especially being able to set NGU to explicitly do direct quadrupling
2) In order to test out NGU-AA I went back to my problem image with original NGU, the one I posted some time ago (480p anime source, soft image with a hand reaching for some cherries). NGU-AA still presents some fairly visible artifacts, especially on the outlines of the fingers. It's bad enough that it's visible while watching normally, when quadrupling 480p. NNEDI3 256 is clearly better in this case still, as the wavy/blotchy artifacts do not appear and the image looks much more natural. I haven't had the time to take screenshots and crop them again - but I'm sure you can reproduce the results based on the original I posted a while back. It's the same style of artifacts as normal NGU produces, but less severe.
3) I've done a simple 1080p > 2160p test, NGU-AA (vhigh) takes ~32ms while NNEDI3 256 takes 58ms (so unusable in real time), without touching any of the other settings. In my 480p config I find that NNEDI3 128, then 64 quadrupling is usable in real time and provides better quality that NGU-AA on that particular source I mentioned.
4) I haven't looked too closely at chroma settings, but generally speaking NGU-AA is faster and I have not noticed any image quality issues due to any NGU flavor here (those are restricted to NGU luma quadrupling).

[Arm3nian]

Originally Posted by madshi

That's all correct, but what does that have to do with 8.0 or 16.0 which you mentioned?

I'm not sure how accurate saying 8.0 was. Just for curiosity, if my source was 1080, and I doubled to 4k, applied 4.0 sharpen edges, then doubled again to 8k, and applied 4.0 sharpen edges, would it or would it not be the same as directly quadrupling 1080 to 8k, and applying 8.0 (if available) sharpen edges?

I just wanted more sharpen edges to be available, and it is, by doubling twice, but wanted the same with direct quadruple. I find the extra sharpness useful when zoomed in on a decent quality decent resolution frame/image. Not that big of a deal, I can use sharpen edges in image enhancements, though not to the same potential.

[har3inger]

Originally Posted by madshi

Which NGU quality level are you using? Which GPU do you have? Generally, the AR speed difference should not be that dramatic for GPUs with decent speed.

Anyway, you can always use the "use downscaling settings" option, then you have full control.

Sharp/High. Yeah, the GPU is a laptop one (AMD 8870m) and is decidedly kinda slow by today's standards. Still, works well enough as an HTPC via hdmi due to pretty low fan noise.

Yes, I'll be using "use downscaling settings" for full control. Makes the most sense.

[madshi]

Originally Posted by Neet009

2. According to my test, NGU AA very high beats NNEDI3 256 as a whole, especially for low-res aliased anime content. NGU AA makes line more focus and less aliasing as expected.

3. I found that NGU AA very high is slightly faster than NNEDI3 64 when zooming 200% without doing any upsaling or downscaling after luma doubling. However, when doing upsaling or downscaling after luma doubling, NGU AA very high is slightly slower than NNEDI3 64. I don't know why but it doesn't matter. I'll replace NNEDI3 with NGU AA on my settings.

4. For chroma upscaling, it's hard to tell the quality differents between NNEDI3 and NGU AA, but I can confirm that NGU AA med is faster than NNEDI3 16.

Thanks!

Originally Posted by Xaurus

I can easily reproduce it, question is, how can I assist you with data?

I'm not sure how I can fix it without being able to reproduce it myself, unfortunately. Is it just a statistics glitch in the OSD? Or does it actually affect playback quality/smoothness? If it's just a statistics glitch, I suppose I'd rather ignore it for the time being, seeing that I can't seem to reproduce it on my PC.

Originally Posted by Bloodred

2)[/B] In order to test out NGU-AA I went back to my problem image with original NGU, the one I posted some time ago (480p anime source, soft image with a hand reaching for some cherries). NGU-AA still presents some fairly visible artifacts, especially on the outlines of the fingers. It's bad enough that it's visible while watching normally, when quadrupling 480p. NNEDI3 256 is clearly better in this case still, as the wavy/blotchy artifacts do not appear and the image looks much more natural. I haven't had the time to take screenshots and crop them again - but I'm sure you can reproduce the results based on the original I posted a while back. It's the same style of artifacts as normal NGU produces, but less severe.

I've tested with that image again, and here are my findings:

Lanczos and even Jinc produce exactly the same artifacts (and Jinc is known to be the most natural and aliasing-free linear scaler!). I think that says a lot. The artifacts do seem to come from the source, considering that even Jinc shows them.

However, NNEDI3 does seem to be able to smooth them out. So my judgement right now is that for extremely soft and wide edges, NNEDI3 manages to apply some Anti-Aliasing, too, while NGU AA is "only" on the same level as Jinc with such edges.

Do you agree with this? If so, it means that NGU AA does not introduce any artifacts on its own, it just has lower AA capability than NNEDI3 for extremely soft & wide edges. That's a very important finding, IMHO, because if NGU AA introduced such kind of artifacts (even if the source doesn't have them!), that would be a major problem. A lower than expected AA capability in some cases is of course not a nice finding, but it's a much less dramatic problem than introducing new weird artifacts into the image.

So now the question would be how you would rate NGU AA vs NNEDI3 with a broader set of sources, not just this one image. Which do you prefer then?

Originally Posted by Arm3nian

I'm not sure how accurate saying 8.0 was. Just for curiosity, if my source was 1080, and I doubled to 4k, applied 4.0 sharpen edges, then doubled again to 8k, and applied 4.0 sharpen edges, would it or would it not be the same as directly quadrupling 1080 to 8k, and applying 8.0 (if available) sharpen edges?

I just wanted more sharpen edges to be available, and it is, by doubling twice, but wanted the same with direct quadruple. I find the extra sharpness useful when zoomed in on a decent quality decent resolution frame/image. Not that big of a deal, I can use sharpen edges in image enhancements, though not to the same potential.

Ah, now I see that your 4.0 and 8.0 numbers were meant to be sharpening strengths? That wasn't at all clear before. I thought you were talking about 4.0 and 8.0 scaling factors or something.

The sharpening after the first doubling has much more effect than sharpening after quadrupling. So if you don't sharpen after doubling but only after quadrupling, just doubling the sharpening strength won't achieve the same effect. However, increasing the sharpening strength even more might introduce sharpening artifacts. The key difference is that sharpening algos look at the pixel neighborhood of a specific size. If you sharpen after doubling, each pixel covers more of the original image than after quadrupling. So practically the pixel neighborhood when sharpening after doubling is bigger than when sharpening after quadrupling. Which explains why sharpening after doubling is more effective. However, using a larger neighborhood also bloats more.

Anyway, a potential solution to this problem might be to modify the sharpening algorithm itself to use a bigger pixel neighborhood when sharpening after quadrupling (instead of doubling). However, that would make the sharpening slower.

I think it would be better to pick a sharper upscaler than to try to sharpen the image. If you like sharp, use "NGU Sharp" and get rid of the extra sharpening algos. If that's still not sharp enough for your taste, then try sharpening the image before upscaling it (see image enhancements).

Originally Posted by har3inger

Sharp/High. Yeah, the GPU is a laptop one (AMD 8870m) and is decidedly kinda slow by today's standards. Still, works well enough as an HTPC via hdmi due to pretty low fan noise.

Yes, I'll be using "use downscaling settings" for full control. Makes the most sense.

JFMI: What is the difference in render times when you switch from Medium to High NGU? And what is the difference when you switch from strict to relaxed AR downscaling?

And can you see a difference in sharpness, when using strict vs relaxed AR? In some images I've found it to be quite noticeable.

[Xaurus]

Originally Posted by madshi

I'm not sure how I can fix it without being able to reproduce it myself, unfortunately. Is it just a statistics glitch in the OSD? Or does it actually affect playback quality/smoothness? If it's just a statistics glitch, I suppose I'd rather ignore it for the time being, seeing that I can't seem to reproduce it on my PC.

Yes exactly, it shows a false rendering time in the OSD. Playback is not affected, however the point is that it can show higher rendering times than what the frame time is - thus one would expect frame drops, but there are none to be found.

In order to have the "proper" rendering time shown, I am forced to change the algorithm for chroma, and then back to what it was.

Some other person said he also saw this bug. Would be excellent to get rid of it since it's quite confusing.

[-sTi-]

I have some problems with assigning keyboard shortcuts in the latest version(s).
I want to do the following: assign a key to get Softcubic 70 luma upscaling (for low quality sources) and one for Lanczos3 AR (for HQ sources; my Intel HD 4600 cannot handle anything more advanced).
This worked in the madVR versions before the redesigned luma scaling options, but now assigning "softcubic" does not work any more . The available choice for keyboard shortcuts is just "cubic" without further choice. The strange thing is that at first the "cubic" shortcut worked but only produced the "bicubic" variant. After I manually changed the dropdown in the luma scaling options to "softcubic" the key I assigned to "cubic" is ignored when pressed.

What can I do to get the desired "softcubic" shortcut?

[FreeFall]

madshi,

Is the new upscaling settings page fine now?

It would be nice if we could choose what to do with chroma when doubling. For example being able to disable chroma doubling and simply select an upscaler to use, or to be able to select what algorithm to double chroma with (NGU Low, Med, High, Very High / Jinc AR, Lancos AR / NNEDI3 16-256 neurons ect).

Probably overkill but I would have liked to have tried out NGU Anti-Alias Very High for both luma and chroma, the current presets are too limited for my use.


When disregarding speed, how would you rate NGU Anti-Alias Very High vs. NNEDI3 256-taps, for image upscaling?

NGU Anti-Alias is pretty close to NNEDI3 now, you've done a great job with it, depending on the source it can be a little too aggressive at removing aliasing which can produce images that have a flat digital look to them. Overall I still prefer NNEDI3 because it's more balanced, keeps more of the image noise / detail intact.


When taking speed into account, how would you rate NGU Anti-Alias vs NNEDI3 at the various quality levels, for image upscaling?

NGU Anti-Alias is almost twice as fast, image doubling 1.2x (quadrupling disabled) PAL interlaced DVD upscaled to 1080p (Geforce GTX 1070) NGU Anti-Alias Very High preset (luma and chroma) ~8.8ms vs NNEDI3 256 neurons / very high preset chroma ~16ms.

I vote to keep NNEDI3 as I don't think NGU Anti-Alias can replace it just yet, highly recommended as a faster alternative though.


Cheers.

[Paul Tronc]

Hello,

I can switch from GTX960 to R9 380 for free. I know that the 380 has an hybrid hevc decoder, but I don't know the real impact on my setup. I mainly read H264 1080p blurays, but I have started encoding my movie stock to HEVC.

Is it worth switching to R9 380? I have a 1080p projector, my render time is globally ok but I use 60Hz frame interpolation so sometimes the render time is borderline. The frame interpolation is managed by the CPU so it doesn't interfer with madvr AFAIK.

[Schwartz]

When selecting doubling instead of scaling, the option 'any (more) scaling' is pretty limited while this is effectively still an important option. It's missing Catmull, Spline, Mitchell. The last two are particularly useful, Mitchell for SD and Spline for general-purpose when Jinc or AR are too expensive. Most of the stuff listed is AR, so choices are effectively made for us. It's not terrible, but I much preferred the old behaviour when it was still a toolbox. Anyway.. I can only repeat that same sentiment so many times so I'll tip my hat now.

[huhn]

Originally Posted by Paul Tronc

Hello,

I can switch from GTX960 to R9 380 for free. I know that the 380 has an hybrid hevc decoder, but I don't know the real impact on my setup. I mainly read H264 1080p blurays, but I have started encoding my movie stock to HEVC.

Is it worth switching to R9 380? I have a 1080p projector, my render time is globally ok but I use 60Hz frame interpolation so sometimes the render time is borderline. The frame interpolation is managed by the CPU so it doesn't interfer with madvr AFAIK.

hybrid decoder is the last thing you want when using madVR.

i don't know which card is faster for madVR but the difference is not going to be a big one.

[-sTi-]

I have a Laptop connected via HDMI to a Panasonic Plasma TV. The dual screen is set up as "extend" (not "duplicate") in Windows 10. I have set up the Laptop to use the NVIDIA 950m instead of integrated Intel graphics when using MPC-HC.
When the refresh rate of the TV is set as 50Hz in Windows graphics control panel everything works fine (the Laptop screen stays at its native 60Hz and MPC-HC running on the TV plays at 50Hz). However, as I use smooth motion, I would like to have the TV run at 60Hz (IIRC this was recommended earlier in this thread).
The problem is I can change this setting in Windows and the playback then starts fine with MPC-HC and madVR in full screen mode (also the correct 60Hz setting is displayed), but if I then e.g. press the right mouse button during the movie the whole TV screen goes black and the video refuses to play after this. Not good. But it gets even worse: If I then try to exit full screen mode with "Alt+Enter", also the Laptop Screen goes black and I can't see anything any more.
Does anyone have an idea what could cause this strange behavior?

[madshi]

Originally Posted by Xaurus

Yes exactly, it shows a false rendering time in the OSD. Playback is not affected, however the point is that it can show higher rendering times than what the frame time is - thus one would expect frame drops, but there are none to be found.

In order to have the "proper" rendering time shown, I am forced to change the algorithm for chroma, and then back to what it was.

Some other person said he also saw this bug. Would be excellent to get rid of it since it's quite confusing.

Well, there's a "secret" probably only devs know: Sometimes you can add a new feature in 2 hours, but sometimes it takes 2 days to find a simple bug. So would you rather want me spending my time chasing an elusive bug which only affects a measurement number in the OSD, and which I can't reproduce on my own PC (which makes it extra hard to fix)? Or would you rather have me add new features like direct quadrupling NGU AA, or NGU Soft/Standard instead? If I asked users for a vote, I'm pretty sure what the majority would want.

Originally Posted by -sTi-

I have some problems with assigning keyboard shortcuts in the latest version(s).
I want to do the following: assign a key to get Softcubic 70 luma upscaling (for low quality sources) and one for Lanczos3 AR (for HQ sources; my Intel HD 4600 cannot handle anything more advanced).
This worked in the madVR versions before the redesigned luma scaling options, but now assigning "softcubic" does not work any more . The available choice for keyboard shortcuts is just "cubic" without further choice. The strange thing is that at first the "cubic" shortcut worked but only produced the "bicubic" variant. After I manually changed the dropdown in the luma scaling options to "softcubic" the key I assigned to "cubic" is ignored when pressed.

Sounds like a bug. I'll check if I can reproduce it.

Originally Posted by FreeFall

Is the new upscaling settings page fine now?

It would be nice if we could choose what to do with chroma when doubling. For example being able to disable chroma doubling and simply select an upscaler to use, or to be able to select what algorithm to double chroma with (NGU Low, Med, High, Very High / Jinc AR, Lancos AR / NNEDI3 16-256 neurons ect).

Probably overkill but I would have liked to have tried out NGU Anti-Alias Very High for both luma and chroma, the current presets are too limited for my use.

I understand, but the reason why I limited the chroma options is because it was very evident from reading the madVR thread that a very high number of users was running madVR with very bad settings (too high chroma settings, too low luma settings). So I've made this decision to "help" users, especially noobs, to use sane settings. Of course I could change the dropdown box to include all the various algorithms, but you can bet that many users would start using crazy configurations again. Furthermore, all the available algorithms could be used either with or without AR, without or without sigmoidal light. So the dropbox would get crazy long.

Originally Posted by FreeFall

When disregarding speed, how would you rate NGU Anti-Alias Very High vs. NNEDI3 256-taps, for image upscaling?

NGU Anti-Alias is pretty close to NNEDI3 now, you've done a great job with it, depending on the source it can be a little too aggressive at removing aliasing which can produce images that have a flat digital look to them. Overall I still prefer NNEDI3 because it's more balanced, keeps more of the image noise / detail intact.

Can you upload a couple of video frames (just a PNG file should suffice) where you prefer NNEDI3? I only need the original (unscaled) image.

Originally Posted by FreeFall

I vote to keep NNEDI3 as I don't think NGU Anti-Alias can replace it just yet, highly recommended as a faster alternative though.

How about chroma upscaling. Can I remove NNEDI3 there, at least?

Originally Posted by Paul Tronc

Is it worth switching to R9 380?

I don't know. If the R9 380 is noticeably faster in games, it will probably also be noticeably faster with madVR. However, latest AMD drivers on Windows 10 seem rather unstable, from the reports I'm seeing here in this thread.

Originally Posted by Schwartz

When selecting doubling instead of scaling, the option 'any (more) scaling' is pretty limited while this is effectively still an important option. It's missing Catmull, Spline, Mitchell. The last two are particularly useful, Mitchell for SD and Spline for general-purpose when Jinc or AR are too expensive.

As usual, I'm willing to reconsider, if you can show me screenshots that prove your point. I highly doubt that after doubling Catmull, Spline or Mitchell would be any better than Bicubic60. You can create comparison screenshots by doubling first, then save a screenshot with exactly 2.0x scaling factor. Then load the screenshot PNG into the media player and turn off doubling. That way you can use Catmull, Spline or Mitchell after doubling.

Originally Posted by -sTi-

I have a Laptop connected via HDMI to a Panasonic Plasma TV. The dual screen is set up as "extend" (not "duplicate") in Windows 10. I have set up the Laptop to use the NVIDIA 950m instead of integrated Intel graphics when using MPC-HC.
When the refresh rate of the TV is set as 50Hz in Windows graphics control panel everything works fine (the Laptop screen stays at its native 60Hz and MPC-HC running on the TV plays at 50Hz). However, as I use smooth motion, I would like to have the TV run at 60Hz (IIRC this was recommended earlier in this thread).
The problem is I can change this setting in Windows and the playback then starts fine with MPC-HC and madVR in full screen mode (also the correct 60Hz setting is displayed), but if I then e.g. press the right mouse button during the movie the whole TV screen goes black and the video refuses to play after this. Not good. But it gets even worse: If I then try to exit full screen mode with "Alt+Enter", also the Laptop Screen goes black and I can't see anything any more.
Does anyone have an idea what could cause this strange behavior?

Sounds like a GPU driver issue. You could try turning off fullscreen exclusive mode. That will probably fix the problem. However, you will need to make sure that playback is still perfectly and reliably smooth afterwards. Normally, fullscreen exclusive mode is the most reliable mode for smooth playback.

[Xaurus]

Originally Posted by madshi

Well, there's a "secret" probably only devs know: Sometimes you can add a new feature in 2 hours, but sometimes it takes 2 days to find a simple bug. So would you rather want me spending my time chasing an elusive bug which only affects a measurement number in the OSD, and which I can't reproduce on my own PC (which makes it extra hard to fix)? Or would you rather have me add new features like direct quadrupling NGU AA, or NGU Soft/Standard instead? If I asked users for a vote, I'm pretty sure what the majority would want.

madshi, please spend your time on what you think is best for madvr. Since it seems this bug was relatively unknown, I am happy to have brought it to the surface. If it will be fixed sooner, that's great, if it will be fixed later, that's great too.

[omarank]

Originally Posted by madshi

* added NGU Anti-Alias algorithm, intended as a potential NNEDI3 replacement

madshi, although NGU Anti-Alias is a definite improvement over other NGU pixart versions, my eyes still seem to like NNEDI3 better. Please see this screenshot comparison.

I think you should not drop a good upscaler like NNEDI3. Before NGU came, NNEDI3 was unarguably the best image doubler in madVR in terms of image quality. People have been using it for several years without any complaint about quality. Even if NGU Anti-Alias was miles better than NNEDI3, still you shouldn’t be replacing it. NNEDI3 may be considered as a good fallback option for image doubling, just like Random dithering is in the choice of dithering algorithms.

[aufkrawall]

I don't manage to see why the NNEDI3 result would be better.
With my cartoon example, it's very obvious that NNEDI3 suffers the problem of connecting lines which shouldn't be connected, unlike NGU AA.

[Bloodred]

Originally Posted by madshi

I've tested with that image again, and here are my findings:

Lanczos and even Jinc produce exactly the same artifacts (and Jinc is known to be the most natural and aliasing-free linear scaler!). I think that says a lot. The artifacts do seem to come from the source, considering that even Jinc shows them.

However, NNEDI3 does seem to be able to smooth them out. So my judgement right now is that for extremely soft and wide edges, NNEDI3 manages to apply some Anti-Aliasing, too, while NGU AA is "only" on the same level as Jinc with such edges.

Do you agree with this? If so, it means that NGU AA does not introduce any artifacts on its own, it just has lower AA capability than NNEDI3 for extremely soft & wide edges. That's a very important finding, IMHO, because if NGU AA introduced such kind of artifacts (even if the source doesn't have them!), that would be a major problem. A lower than expected AA capability in some cases is of course not a nice finding, but it's a much less dramatic problem than introducing new weird artifacts into the image.

I hadn't even thought to try Jinc, but yes, after having a look at the image it produces I agree. NGU AA doesn't introduce the artifacts by itself. As you've said though, NNEDI3 does a better job with on this source, but NGU AA doesn't add anything bad to the image by itself, which is what I thought at first. NNEDI3 256+256 is quite clearly better, but too slow by far for realtime usage. NNEDI3 128+64 however is slightly faster than NGU AA vhigh (26ms vs. 28ms) and still provides a better image.

Originally Posted by madshi

So now the question would be how you would rate NGU AA vs NNEDI3 with a broader set of sources, not just this one image. Which do you prefer then?

I've tried it out with other 480p and 720p sources (no 1080p because that's entirely NGU Sharp territory for me, haven't been disappointed yet). This also means that it was always quadrupling. Overall I'd say NGU AA is a good replacement for NNEDI3. For the most part I can get equivalent image quality if not slightly better for the same render time. I imagine if you add direct quadrupling for NGU AA the performance advantage for NGU AA will only become much higher - which is nice, that translates to less fan noise. In general I'd say that NGU AA appears to be slightly sharper than NNEDI3, which I prefer for all the sources I've tried other than the one mentioned above.

NGU AA is overall better than NNEDI3 I think, but not universally better. Also I'd love to see direct quadrupling for NGU AA. Direct quadrupling 720p content with NGU Sharp is pretty great!

[madshi]

Originally Posted by omarank

madshi, although NGU Anti-Alias is a definite improvement over other NGU pixart versions, my eyes still seem to like NNEDI3 better. Please see this screenshot comparison.

Did you compare with a number of different sources? All other users so far seem to prefer NGU AA *overall*, although NNEDI3 might win with some sources.

FWIW, in your comparison image, some of the lines are better anti-aliased by NNEDI3, but the dress of the girl is clearly sharper in the NGU AA image.

Originally Posted by omarank

I think you should not drop a good upscaler like NNEDI3. Before NGU came, NNEDI3 was unarguably the best image doubler in madVR in terms of image quality. People have been using it for several years without any complaint about quality. Even if NGU Anti-Alias was miles better than NNEDI3, still you shouldn’t be replacing it. NNEDI3 may be considered as a good fallback option for image doubling, just like Random dithering is in the choice of dithering algorithms.

The thing is that many users have stability problems with NNEDI3/OpenCL. Problems range from graphical corruption, over Windows 10 updates removing OpenCL support up to crashes. Also, if I could remove NNEDI3, I could get rid of all the ugly OpenCL code.

Anyway, I won't remove NNEDI3 immediately, if there are several users opposed to it.

I would like to thin out the options dialog, though, so I'd still like to know if anybody were shocked if I remove NNEDI3 as a *chroma* upscaling algorithm? The list is too long for my taste, and considering that chroma upscaling is less visible than luma, and most users seem to prefer NGU AA over NNEDI3, it shouldn't do too much harm?

Originally Posted by aufkrawall

With my cartoon example, it's very obvious that NNEDI3 suffers the problem of connecting lines which shouldn't be connected, unlike NGU AA.

Yes. From my internal list of test images, 3 are very clear wins for NGU AA, one is a clear win for NNEDI3, and all others are a wash, with maybe NGU AA being slightly better in most of them. To me it seems NGU AA wins overall, although NNEDI3 might win in some cases.

Originally Posted by Bloodred

I hadn't even thought to try Jinc, but yes, after having a look at the image it produces I agree. NGU AA doesn't introduce the artifacts by itself. As you've said though, NNEDI3 does a better job with on this source, but NGU AA doesn't add anything bad to the image by itself, which is what I thought at first.

Same here. Glad we cleared that up. And I'm happy that NGU AA is not responsible for adding those weird artifacts.

Originally Posted by Bloodred

NNEDI3 256+256 is quite clearly better, but too slow by far for realtime usage. NNEDI3 128+64 however is slightly faster than NGU AA vhigh (26ms vs. 28ms) and still provides a better image.

With this specific image, yes. It's the only one in my test set where NNEDI3 clearly wins.

Originally Posted by Bloodred

I've tried it out with other 480p and 720p sources (no 1080p because that's entirely NGU Sharp territory for me, haven't been disappointed yet). This also means that it was always quadrupling. Overall I'd say NGU AA is a good replacement for NNEDI3. For the most part I can get equivalent image quality if not slightly better for the same render time. I imagine if you add direct quadrupling for NGU AA the performance advantage for NGU AA will only become much higher - which is nice, that translates to less fan noise. In general I'd say that NGU AA appears to be slightly sharper than NNEDI3, which I prefer for all the sources I've tried other than the one mentioned above.

NGU AA is overall better than NNEDI3 I think, but not universally better. Also I'd love to see direct quadrupling for NGU AA. Direct quadrupling 720p content with NGU Sharp is pretty great!

That mostly matches my own findings, but I haven't tested a whole lot of different sources yet, to be honest.