madVR - high quality video renderer (GPU assisted)

[madshi]

I'm curious because I've started using Smooth motion (as I've heard it's great with high refresh rates)
FWIW, if you can set your display's refresh rate to a straight multiply of the source frame rate (e.g. 23.976Hz, 48.952Hz or 71.928Hz for 23.976fps sources) then you should probably disable smooth motion FRC. The real purpose of smooth motion FRC is to remove pulldown judder when movie framerate and display refresh rate don't match. Ok, in theory you can also use smooth motion FRC as a Reclock replacement to get endless playback without frame drops, and for that using a very high refresh rate is very useful. But I still believe it's better to that the display to a multiply of the source framerate and to disable smooth motion FRC. You might get a frame drop/repeat every couple of hours, but nobody has died from that yet (well, as far as I know)...

I think you just did too much at the same time, when you introduced too many options in one of the last updates, that's why people really get confused and try out 2 million different settings and they never really provide meaningful feedback for only one option.
Probably, but I would have felt bad, "hiding" already existing features just to get better feedback.

When we did the dithering tests, there was great communication, screenshots and lots of example shots where you could very clearly detect the positives and negatives and you also reacted very fast to improve them. With finesharp, this was not the case, since no one other than TheLion, me and you even provided screenshots. That's 3 people out of maybe 20 that basically just said "I like A better than B", but they could not tell us, why.
True, but the way things were going I didn't expect the feedback situation to change much, so I made the decision based on the feedback I got. If I had known that 2 days later suddenly contradicting feedback would show up, obviously I would have delayed my decision. Anyway, no harm done yet. Easy enough to revert my decision. Which doesn't mean I will, though. Not decided yet...

One of the main reasons why a lot of DVDs were so bad, was not because of their low native resolution, but because of their excessive ringing that was applied in post. They are almost completely unwatchable when you upscale them to HD or 4K. Finesharp with the current LL will amplify that even more and that doesn't make any sense when you can just use finesharp with no LL without these negatives.
But in the left half of the squirrel test pattern, actually there were less ringing artifacts with LL turned on compared to off. So I'm not sure if the situation is that simple as you make it sound.

It's definitely LL that makes contoures look too thick and dark with FS.
not LL:
http://abload.de/thumb/nollzjpua.png (http://abload.de/image.php?img=nollzjpua.png)

LL:
http://abload.de/thumb/llhmqr0.png (http://abload.de/image.php?img=llhmqr0.png)

Sample (1 frame):
http://www52.zippyshare.com/v/lTUNZizZ/file.html

So I vote to make not LL the new default option. In fact, I don't see any reason for having a LL option if it distorts images as shown.

LL is devilish. Always, it seems. :devil:

As for thinning: With low values I don't see any effect (at least no positive). Combined with SuperRes, it can even look worse because it already seems to introduce tiny bits of aliasing which then get stronger. With higher values, this aliasing also becomes visible without SuperRes. So I don't see any point in this parameter, thus I vote to set it to 0 by default and probably even remove this option.
Thanks for the sample. I'm not sure I can reproduce the aliasing problem, though. Using super-xbr for doubling, followed by FineSharp in upscaling refinement with rather high thinning setting. Can you post a screenshot (preferably without SuperRes) where the aliasing is visible, and maybe mark it in MS Paint with a red arrow or something? And please let me know which exact upscaling/processing options you're using. Thanks!

"Better" is not objective. LL scaling seems to always manipulate brightness in a way that leads to an inaccurate result compared to the source. At least from what I've seen.
Scientifically it should be the other way round, at least when downscaling. Do you have some samples?

Madshi, are there any knobs to turn for Super-xbr with regards to sharpness outside of SuperRes?
Maybe a different kernel, which would make things softer. Not sure yet, though.

Ok. However, I for myself have already stumbled over three (with FS now four) more or less "real world" examples in which LL was clearly worse.
Well, then I suppose it might be best if madshi leaves us the option to choose, like with scaling.
Samples?

No, there is not going to be a LL option for sharpening, sorry. There are way too many options in madVR already. At some point I'll remove several (many?) of the current options. (Of course I'll ask for feedback first.)

FWIW, I just opened the Settings to confirm the above, and sure enough "Double Chroma Resolution" is checked. That must be the new default setting, as I've never checked that option in the past. I think that maybe "Double Luma Resolution" is also getting checked as a new default (both set to "NNEDI3, 16 neurons"), as I'm pretty sure I didn't check double luma on this laptop (it can't handle it).
There's a bug in the settings dialog which results in "Double Chroma Resolution" to always be checked when you enter the settings dialog with "Double Luma Resolution" checked. But that's just a bug in the settings dialog. Whatever settings you change and store, are saved and used correctly, as far as I can see.

For NEDI and super-xbr you cannot disable "Double Chroma Resolution" if "Double Luma Resolution" is enabled, becase those two algorithms always double both at the same time by design.

I know madshi wants some report about FS first, but it will be more practical to me to make an overview of the lasts features.

Finesharp
I used it years ago when Didée released it on avisynth after an user of doom9 ask him what was the best sharpener for HD source.
So here we have a important point.
Finesharp is definitively not an option to considering if we talk about sources < 720p.
FS is a factory for aliasing, distorsions, and artifacts. He has to be used only with very HQ content. Linear Light or not, he's too strong for low quality sources, the result is always horrible.
At least in my opinion, and my opinion is the only thing i can provide here.

Thinning control seems to be the equivalent setting to "xstr" in the Didée's avisynth script.
In others words, xstr/thinning is the setting than you absolutely want to turn down the most, unless you're an aliasing fan.

Finesharp is a very destructive sharpener, and to be efficient, i think he will be only used with Blu-ray or 1080p content.

And even on these types of sources, i think the strength should not exceed 1.2, in upscaling refinement, and no more than 0.7/0.8 in image enhancement.

Repair was set at 1.0, and to be honest, i don't know what to think about that. By the fact, i have not succeed to rule if repair fonction is equivalent to the "cstr" factor in the avisynth script. I think is some of, but not completely, cos i suspect the madvr version is quite different in some ways than the original script.
For fact, if repair have some strong common point with cstr, so 1.0 seems to be too high for me.
Thanks for your feedback - good 2nd post on the forum! :D

Yes, "thinning" is "xstr". However, "repair" is an algorithm tweak implemented by -Vit- when converting the AviSynth script to GPU pixel shaders. Basically "repair" tries to avoid/undo the aliasing problem that occurs when using "thinning". So your experience with the AviSynth script introducing aliasing when using "thinning" needs to be re-evaluated. There is still some aliasing, but it's less a problem compared to the AviSynth script. I'm still playing with the proper "repair" value. The "cstr" option is set automatically by madVR, depending on which strength you've selected.

LumaSharpen

I discovered this sharpener with a previous build of madVR.
I think it's an adaptation of a SweetFx option, and i welcome it with a sort of mistrust, because i hate SweetFX :d
Anyway, i tried to let him a chance, and after some tests, i have to say he can have some utility, on very HQ content, again.
With really moderate settings, and considering is just an eye candy toy, he adds some kind of pop effect wich is not unpleasant.

As he active on the luma, he brighter just some objects and pixels in a frame, and in some ways, i like it.

Maybe a lead for futures evolutions, add an usharp mask (like Unsharp HQ) to work with luma sharpen will open a new way for those who wants to build a very pop picture. If madshi let Luma Sharpen in madVR, and i'll like he will, this will be an option to hardly considerate.

An other opinion is if you're upscaling the content, you should definitively not turn on a sharpener in image enhancement, whatever it is.
I'm not a big fan of LumaSharpen. It is definitely quite fast, that's its big advantage. But it still introduces quite noticable ringing. At some point I'm probably going to replace LumaSharpen with a slower alternative, something similar to LimitedSharpenFaster, which doesn't ring as much. But for now I'll probably keep LumaSharpen because it's so nicely fast.

Super-xbr

As i play almost only 1080p content on 3440x1440 monitor and 4K VP, and as i have the rigs to motorize anything, i was a faithful user of NNEDI3 in chroma upscaling and Double resolution.
But super-xbr have insinuate the doubt in me.
On some aspects, super-xbr outclasses NNEDI. Which are aliasing control, clarity and natural look. But he's beaten by NNEDI3 on cleanliness, ringing control and precision.
This is worth for almost all types of content.

Then, super-xbr is incredibly less greedy for a result which is different but fairly comparable in terms of quality to NNEDI3, event at 128 or 256 neurons.

And more than that, it seems to me than super-xbr produce fantastic result in chroma upsampling.
I'm glad you like it that much! To be honest, I was a bit disappointed by the feedback so far. Most users didn't seem to like it much. Probably my fault, due to choosing test images which already had a lot of ringing in them, so it looked like super-xbr would ring a lot.

1080p content on 1080p TV
- Custom Res at 2560x1440 in Nvidia panel Control (no DSR)
- Chroma upscaling : super-xbr (without SuperRes)
- Luma/chroma doubling : super-xbr
- Upscaling refinement : Super Res (NNEDI3 defaults, 4 passes), FS (strength 0.4, thinning 0.007), LS (strength : 0.25, clamp : 0.045, radius 0.3)
- image downscaling : C-R AR LL
Who's doing the downscaling here? The GPU driver, I suppose? Does this really look better than straight 1080p playback? Oh well, SuperRes doesn't activate for 1:1 playback, maybe that's the reason?

[RainyDog]

I'm not a big fan of LumaSharpen. It is definitely quite fast, that's its big advantage. But it still introduces quite noticable ringing. At some point I'm probably going to replace LumaSharpen with a slower alternative, something similar to LimitedSharpenFaster, which doesn't ring as much. But for now I'll probably keep LumaSharpen because it's so nicely fast.

Have you looked at Adaptive Sharpen (http://forum.doom9.org/showthread.php?t=172131) madshi?

[iSunrise]

...Scientifically it should be the other way round, at least when downscaling. Do you have some samples?
While that may be true, the real-world results with madVR are the opposite though, thats why LL was only recommended together with Catmull-Rom downscaling. At least that's what we concluded after we did several tests years ago and e.x. 6233638 and cyberbeing were two people that also strictly recommended against it. LL on some downscalers was full of artefacts compared to gamma light.

[har3inger]

I'm glad you like it that much! To be honest, I was a bit disappointed by the feedback so far. Most users didn't seem to like it much. Probably my fault, due to choosing test images which already had a lot of ringing in them, so it looked like super-xbr would ring a lot.



Well, here's another vote saying super-xbr is a lot better than NEDI and a fair bit better than jinc3ar for luma. Maybe it's because it's the first doubling algorithm I've ever tried (nnedi3 doesn't work on my laptop), but it looks extremely clean and less granular compared to jinc3ar. I haven't really noticed any substantial ringing or haloing that wasn't there with jinc3ar already. It's my new default scaler for everything now.

I can't comment on super-xbr for the initial chroma upscale, however. In principle I don't feel great about using "smart" scalers for chroma, because the "other artifacts" it can introduce might create mismatches with the unscaled luma channel, which is expecting a conventionally upscaled chroma. This might just be a fear founded on ignorance though.

I dislike NEDI mostly because it likes to destroy small text (0.5-1.5 pixel thickness letters) in credits and title screens on lower quality sources where doubling/quadrupling is most useful. The algorithm somehow makes letters lumpy and can introduce a patchiness in their luminance, SuperRes or no.

[Barnahadnagy]

I'm glad you like it that much! To be honest, I was a bit disappointed by the feedback so far. Most users didn't seem to like it much. Probably my fault, due to choosing test images which already had a lot of ringing in them, so it looked like super-xbr would ring a lot.

+1 for S-XBR. Its pretty awesome quality for the performance. You can say NNEDI is better all you want, I still wont be able to run it ;).

[madshi]

Have you looked at Adaptive Sharpen (http://forum.doom9.org/showthread.php?t=172131) madshi?
No, I wasn't aware of that one. I might give it a try next weekend.

While that may be true, the real-world results with madVR are the opposite though, thats why LL was only recommended together with Catmull-Rom downscaling. At least that's what we concluded after we did several tests years ago and e.x. 6233638 and cyberbeing were two people that also strictly recommended against it. LL on some downscalers was full of artefacts compared to gamma light.
Are you talking about:

A) LL adding ringing artifacts? Or:
B) LL changing the brightness of the original source?

- Your recent posts were about A).
- Your screenshots showed A).
- aufkrawall talked about B).
- Your post is a reply to mine, replying to aufkrawall about B).
- 6233638's and cyberbeing's LL scaling recommendations are because of A).
- AFAIR, neither 6233638 nor cyberbeing have ever complained about B).

So far I've not seen a single screenshot yet supporting B). So what is it you're talking about? A) or B)? If it's A) then your post made no sense because aufkrawall and I have been discussing B). If you were now talking about B) then you mentioning 6233638 and cyberbeing made no sense because their scaling recommendations were motivated by A) and had nothing to do with B) at all.

Please don't misunderstand me. I appreciate your feedback very much. Having feedback that differs from other feedback can be helpful to find the best solution. But please don't mix different things together, or use arguments for one thing (A) while talking about a different thing (B). It's very frustrating for me to discuss that way, because simply just explaining how things are mixed up costs a lot of time and serves no real purpose.

If you do think that B) exists then please post some screenshots where LL downscaling modifies the brightness of the source (while not using LL doesn't). But don't use 6233638 and cyberbeing in that discussion then, because referencing them only makes sense when talking about A).

Well, here's another vote saying super-xbr is a lot better than NEDI and a fair bit better than jinc3ar for luma. Maybe it's because it's the first doubling algorithm I've ever tried (nnedi3 doesn't work on my laptop), but it looks extremely clean and less granular compared to jinc3ar. I haven't really noticed any substantial ringing or haloing that wasn't there with jinc3ar already. It's my new default scaler for everything now.

I can't comment on super-xbr for the initial chroma upscale, however. In principle I don't feel great about using "smart" scalers for chroma, because the "other artifacts" it can introduce might create mismatches with the unscaled luma channel, which is expecting a conventionally upscaled chroma. This might just be a fear founded on ignorance though.

I dislike NEDI mostly because it likes to destroy small text (0.5-1.5 pixel thickness letters) in credits and title screens on lower quality sources where doubling/quadrupling is most useful. The algorithm somehow makes letters lumpy and can introduce a patchiness in their luminance, SuperRes or no.
Thanks for your feedback.

+1 for S-XBR. Its pretty awesome quality for the performance. You can say NNEDI is better all you want, I still wont be able to run it ;).
Cool, thanks.

[aufkrawall]

Thanks for the sample. I'm not sure I can reproduce the aliasing problem, though. Using super-xbr for doubling, followed by FineSharp in upscaling refinement with rather high thinning setting. Can you post a screenshot (preferably without SuperRes) where the aliasing is visible, and maybe mark it in MS Paint with a red arrow or something? And please let me know which exact upscaling/processing options you're using. Thanks!

Thinning 0:
http://abload.de/thumb/09wqd3.png (http://abload.de/image.php?img=09wqd3.png)

Thinning 0.05:
http://abload.de/thumb/0.05fboxf.png (http://abload.de/image.php?img=0.05fboxf.png)

Strength: 1
Chroma: Jinc3AR
Luma: NNEDI3 64
SuperRes (mild custom values)
Should be independent from scaling settings, but in doubt I always use NNEDI3 to have the cleanest result.

Well, thinning value of 0.05 is rather high, but apart from strength it'S still doing the same as with 0.02.
So, what is the point of thinning? I can't observe a positive effect and it seems to be potentially harmful.
Unfortunately, I forgot what I did to see the aliasing also with lower values in combination with SuperRes.


Scientifically it should be the other way round, at least when downscaling. Do you have some samples?

As I said, I will try to create one in the next days.
Some pages back I posted an image from Wikipedia and also fonts become harder readable.



No, there is not going to be a LL option for sharpening, sorry. There are way too many options in madVR already. At some point I'll remove several (many?) of the current options. (Of course I'll ask for feedback first.)

You also offer different options for deinterlacing for animation and filmed content, I don't see any difference in the situation...
Btw: Also game-captures profit a lot by FS due to chroma subsampling blur and they often show structures that aren't common and thus LL shouldn't be used with them neither.
You really make FS less usable with such content if you keep the option removed. :(
Already two real world cases...

[toniash]

Have you looked at Adaptive Sharpen (http://forum.doom9.org/showthread.php?t=172131) madshi?

It's very good and ONLY ONE KNOB to adjust!

[StinDaWg]

Super-xbr reminds me of running NNEDI3 with a light sharpen afterwords, like SuperRes. I quite like it.

I assume Super-xbr shares code from SuperRes due to both the look and the pixel shift it creates?

With clean sources it looks quite sharp and nice. I think I may actually prefer it over NNEDI3 for blu-rays and other clean video.

The GPU hit is very minimal compared to NNEDI3 too. :)

Good job!

Now I just need to figure out if super-xbr or NNEDI3+SuperRes (after upscale) produces a better overall picture.

[RyuzakiL]

That is just false. Both LL on and LL off will cause problems in certain content.
How do you think we arrived on the decision for LL on at the first place (and incidentally noone objecting)? Because someone thought LL on looked better? Exactly!



If you run a CrossFire or SLI setup, you better know how to setup a profile to disable it, since its not going to work with everything.
Because if you don't, your in for a world of hurt.


Just an update, making a profile of Madvr on AMD CCC and set the crossfire mode to disabled, doesn't work at all.

I'm Using 2X AMD HD7850

If I use NNEDI3 128 Neurons in Chroma Upscaling crossfire still turns on and I had a lot of dropped frames, i thought AMD Cards runs fine using NNEDI3 compared to NVIDIA's. But when I use Super-XBR, Crossfire doesn't turns on whether i make a profile of Madvr or Not on AMD CCC. This is coming from a fresh clean install of OS + Latest Beta Drivers from AMD.

it seems AMD cards wasn't suited using NNEDI3 though i read somewhere that AMD's R9 270 can run NNEDI3 128 Neurons just fine. So wth? seems to me it's a driver conflict.

So for now I'll just keep on using Super-XBR and hope for further development for this algo., It's lightweight and Picture Quality is on Par with NNEDI3 and Crossfire isn't turning on when using Super-XBR so that's a plus on my book.

[Eyldebrandt]

Thanks for your feedback - good 2nd post on the forum! :D

Yes, "thinning" is "xstr". However, "repair" is an algorithm tweak implemented by -Vit- when converting the AviSynth script to GPU pixel shaders. Basically "repair" tries to avoid/undo the aliasing problem that occurs when using "thinning". So your experience with the AviSynth script introducing aliasing when using "thinning" needs to be re-evaluated. There is still some aliasing, but it's less a problem compared to the AviSynth script. I'm still playing with the proper "repair" value. The "cstr" option is set automatically by madVR, depending on which strength you've selected.

Yah, i've figured it out, because a simple rewrite of cstr with a factor of 1.0 would have result of a ton of blur :d
For that matter,i will let a 2nd chance to thinning, to see where it goes.


I'm not a big fan of LumaSharpen. It is definitely quite fast, that's its big advantage. But it still introduces quite noticable ringing. At some point I'm probably going to replace LumaSharpen with a slower alternative, something similar to LimitedSharpenFaster, which doesn't ring as much. But for now I'll probably keep LumaSharpen because it's so nicely fast.

Imo, I would have seen LSF instead of Finesharp. For the reason I think that lumasharpen and LSF don't do the exacts same things.

But FS had the advantage to be already wrote as a shader.

If you want something like LSF and something fast, maybe the LSF adaptation by Emmanuel Piat could interest you.
It's LimitedSharpenfaster amputee of most of his fonctions to let a simple sharpener in state, very effective, extremely fast, and with high artifacts control.
Let know if that what you search.


I'm glad you like it that much! To be honest, I was a bit disappointed by the feedback so far. Most users didn't seem to like it much. Probably my fault, due to choosing test images which already had a lot of ringing in them, so it looked like super-xbr would ring a lot.

Maybe.
But from what i can note in comparison of NNEDI3 w/e neurons and that in any settings, quality is equivalent, and performance gap is impossible to catch up.
So, to my opinion, unless you find an s-xbr/nnedi3 killer, all in on s-xbr.


Who's doing the downscaling here? The GPU driver, I suppose? Does this really look better than straight 1080p playback? Oh well, SuperRes doesn't activate for 1:1 playback, maybe that's the reason?

Yep, GPU driver.
It's an application with only madVR of the manual supersampling method.
Initially, upscaling (and other process, like deblocking, debanding, sharpening, etc...) was made in avisynth, sometimes with multiples layers, then madVR was downscaling to the final resolution.

Now, willingly or not (not i suspect :d ), the same processing is made possible with only madVR, which is a very better way.

And yes, that look better, it's the exact same process than downsampling in videogame. Less aliasing, a more natural look pictures, more clarity, etc...

[XMonarchY]

I finally found the cause of those 10-15 stutters per 45 minute video - Direct3D 11. It just does not work well for me and I have no idea why... Whether I tick the Sync Every Frame option or not makes 0 difference. Adjust all other settings does nothing.

[PetitDragon]

Yep, GPU driver.
It's an application with only madVR of the manual supersampling method.
Initially, upscaling (and other process, like deblocking, debanding, sharpening, etc...) was made in avisynth, sometimes with multiples layers, then madVR was downscaling to the final resolution.

Now, willingly or not (not i suspect :d ), the same processing is made possible with only madVR, which is a very better way.

And yes, that look better, it's the exact same process than downsampling in videogame. Less aliasing, a more natural look pictures, more clarity, etc...

Haha, finally I see the method as exactly the same as mine.:D I doubt it could be made with only madVR tho, due to very heavy GPU load. :p

[lanzorg]

I'm not a big fan of LumaSharpen. It is definitely quite fast, that's its big advantage. But it still introduces quite noticable ringing. At some point I'm probably going to replace LumaSharpen with a slower alternative, something similar to LimitedSharpenFaster, which doesn't ring as much. But for now I'll probably keep LumaSharpen because it's so nicely fast.


I don't like LumaSharpen and my laptop can't handle FineSharp.
I used LimitedSharpenFaster with Avisynth before and it could be really great to get it inside madVR.

[yukinok25]

Do you see any problems with repair set to 1.0? If so, I'd love to see a screenshot. I was planning to remove the repair option and set it fixed to 1.0. So if you found a situation where repair 1.0 is harmful, it would be good to know.


Actually after playing with the repair setting for a couple of days on different video sources, I have to say that the image looks better at 1.0.
Still mode 3 is the best option for me thus.

At this point I think its good to set the repair fix at 1.0.

Questions out of topic: when I choose DX11 my PC fan ( I7 4700QM 970M GTX) spins like crazy. Do I have to consider it normal?

[AngelGraves13]

I've been using Jinc for Chroma and Image Upscaling for a while now. Is there any reason to use Image Doubling with super-xbr for both chroma and luma? I'm not seeing much of a difference.

Rendering times are below 15 ms on average with Jinc and about 20 ms with super-xbr.

[Warner306]

I've been using Jinc for Chroma and Image Upscaling for a while now. Is there any reason to use Image Doubling with super-xbr for both chroma and luma? I'm not seeing much of a difference.

Rendering times are below 15 ms on average with Jinc and about 20 ms with super-xbr.

Image doubling is only useful for resolutions where the image is doubled (e.g. SD -> 1080p or 1080p -> 2160p). As for chroma upscaling, super-xbr or NNEDI3 should be better, but it will likely be difficult to determine the difference because the chroma layer is inherently soft and difficult to differentiate.

So, the impact of image doubling should be very apparent for scaling factors 2x and larger but less apparent for all other resizes.

[mindz]

I think alot of newbies in video processing like me are a bit lost (or maybe just me?) in when the new processing methods take place, and thus i havent used them - because i dont know what it will do to the picture and when.

I have three questions to clarify it for me.

1. What is the difference between FineSharp, LumaSharpen and SuperRes?
2. When are they applied, one pre scaling and other post scaling?
3. Will the processing methods work when you use DXVA for scaling, or is everything then ignored?

[iSunrise]

...But in the left half of the squirrel test pattern, actually there were less ringing artifacts with LL turned on compared to off. So I'm not sure if the situation is that simple as you make it sound.
Somehow I really begin to think you are looking at different images than I am. With the shots I provided there is no way that anyone could argue that LL has less ringing and less artifacts than no LL. It is not even close. I have even looked at the shots on different screens (even at work) and there's no way in hell that the no LL shot has more ringing than LL.

finesharp with no LL (no additional ringing):
http://abload.de/thumb/sample_thelion_fineshnjp77.png (http://abload.de/image.php?img=sample_thelion_fineshnjp77.png)

finesharp with LL (heavy additional ringing):
http://abload.de/thumb/sample_thelion_fineshjgoi2.png (http://abload.de/image.php?img=sample_thelion_fineshjgoi2.png)

If you refer strictly to the diagonal lines test, please zoom into your own images and show me that there is actually less ringing, because I could not reproduce that. LL always produces darker edges with more heavy ringing. I suspect that's also why you and maybe some others probably prefer the LL look. When you look at the shots above quickly one after another, you can see that this is a very deceiving thing that is happening. The picture might be interpreted to look "better" for some, but the picture is just heavily altered with LL.

Are you talking about:

A) LL adding ringing artifacts? Or:
B) LL changing the brightness of the original source?
I am talking both about A) as well as B), like I already explained. Not only does the finesharp LL add very heavy ringing artefacts (see the shots above) to the image, but it also darkens the image (on the squirrel image above, see the long hair at the left side, close to the image border, it will get darker). Also, I wrote several times that there is a certain "blackness" that consumes some darker image details, just look at all the square borders in the following examples. These are the original shots that I provided again:

finesharp with no LL (no additional ringing, darker picture details are still intact, therefore the picture is obviously brighter (look at the object edges)
http://abload.de/thumb/sample1_finesharp_3_1mept8.png (http://abload.de/image.php?img=sample1_finesharp_3_1mept8.png)

finesharp with LL (heavy additional ringing, darker picture details are crushed, the ringing/borders consume some object edges):
http://abload.de/thumb/sample1_finesharp_3_1jkoon.png (http://abload.de/image.php?img=sample1_finesharp_3_1jkoon.png)

The edges of the squares get completely destroyed by the heavy ringing.

To sum up my findings:
The higher the upscaling factor (the more upscaling is needed), the heavier the added ringing artefacts of finesharp LL become and they completely destroy image detail with it. Even if you set repair to 1.00 (see above). Possible solutions: Either find a way to remove the heavy ringing with finesharp LL, which doesn't destroy picture detail or please consider changing the defaults.

And if you really insist I can do countless other examples, which show exactly the same kind of problems, but that will have to wait for the weekend.

[Braum]

I think alot of newbies in video processing like me are a bit lost (or maybe just me?) in when the new processing methods take place, and thus i havent used them - because i dont know what it will do to the picture and when.

I have three questions to clarify it for me.

1. What is the difference between FineSharp, LumaSharpen and SuperRes?
2. When are they applied, one pre scaling and other post scaling?
3. Will the processing methods work when you use DXVA for scaling, or is everything then ignored?

Check this thread out, it's very useful and you'll find the answer to (most of) your questions ;)

Madvr options explained by Asmodian
https://www.doom9.org/showthread.php?t=171787

1 : FineSharp: A sharpener originally by Didée that attempts to keep local energy close to the original.

LumaSharpen: Blurs the original pixel with the surrounding pixels and then subtracts the blur.

SuperRes: This is a post process method. From Shiandow
Calculate an initial guess (using the configured upscaler)
Downscale and calculate differences with original image.
Scale those differences to the final size (medium = SoftCubic50, high = Jinc3)
Improve guess by:
Softening the image
Subtracting differences with the original image
Sharpening (optional)
Removing aliasing (optional)
Removing ringing (optional)
Repeat steps 2-4 several times.


2 : If you activate Finesharp and Lumasharpen in image enhancements then it's applied before any scaling.
If you activate Finesharp, Lumasharpen and SuperRes in upscaling refinement, then it's applied after chroma upscaling, image downscaling, image doubling and image upscaling. Upscaling must be active (you can check with crtl+J) in order to have upscaling refinement to take effect.

3 : If you use DXVA and there is no need for scaling then only image enhancements is applied.
If you use DXVA and there is a need for scaling, then image enhancements and upscaling refinement are applied.

Sorry for my english.

[madshi]

Thinning 0:

Thinning 0.05:
[url=http://abload.de/image.php?img=0.05fboxf.png]

Strength: 1
Chroma: Jinc3AR
Luma: NNEDI3 64
SuperRes (mild custom values)
Should be independent from scaling settings, but in doubt I always use NNEDI3 to have the cleanest result.

Well, thinning value of 0.05 is rather high, but apart from strength it'S still doing the same as with 0.02.
So, what is the point of thinning? I can't observe a positive effect and it seems to be potentially harmful.
Unfortunately, I forgot what I did to see the aliasing also with lower values in combination with SuperRes.
Ok, thanks. Is it possible you were testing with v0.88.11 originally and didn't have repair set to 1.0? That might explain why you saw aliasing problems with lower thinning values. I think I need to set repair even higher than 1.0 when using thinning values as high as 0.05. That might get rid of the aliasing.

Thinning can be very useful, but it depends on the source and whether you apply it before or after upscaling etc.

You also offer different options for deinterlacing for animation and filmed content, I don't see any difference in the situation...
Btw: Also game-captures profit a lot by FS due to chroma subsampling blur and they often show structures that aren't common and thus LL shouldn't be used with them neither.
You really make FS less usable with such content if you keep the option removed. :(
Already two real world cases...
In the long run there is not going to be an option, but that doesn't mean I've already fully decided yet which way to set the linear light option. I might force it off instead of on, or try middle ground. Let's wait and see what the discussion will bring.

I assume Super-xbr shares code from SuperRes due to both the look and the pixel shift it creates?
Nope, it's got nothing to do with SuperRes. Neither SuperRes nor super-xbr have a pixel shift, btw. The only algos which have a pixel shift are NEDI and NNEDI3. madVR fixes the NEDI/NNEDI3 pixel shift when you activate SuperRes, though.

For that matter,i will let a 2nd chance to thinning, to see where it goes.
Don't use too high values for now, though, I still need to fine tune the repair parameter. It seems to need higher repair values if you use higher thinning values.

Imo, I would have seen LSF instead of Finesharp. For the reason I think that lumasharpen and LSF don't do the exacts same things.

But FS had the advantage to be already wrote as a shader.

If you want something like LSF and something fast, maybe the LSF adaptation by Emmanuel Piat could interest you.
It's LimitedSharpenfaster amputee of most of his fonctions to let a simple sharpener in state, very effective, extremely fast, and with high artifacts control.
Let know if that what you search.
Sounds good to me, but I suppose it's not available as a HLSL pixel shader yet? Where can I find that adaptation, in any case?

But from what i can note in comparison of NNEDI3 w/e neurons and that in any settings, quality is equivalent, and performance gap is impossible to catch up.
So, to my opinion, unless you find an s-xbr/nnedi3 killer, all in on s-xbr.
Well, NNEDI3 does handle sources with lots of ringing in them better than super-xbr, and NNEDI3 manages to tighten up lines/features, which super-xbr doesn't do. But I do like super-xbr, and I think it has a very good performance vs quality ratio.

Questions out of topic: when I choose DX11 my PC fan ( I7 4700QM 970M GTX) spins like crazy. Do I have to consider it normal?
Have you tried the tips in this post?

http://forum.doom9.org/showthread.php?p=1724688&highlight=problems#post1724688 (http://abload.de/image.php?img=09wqd3.png)

Somehow I really begin to think you are looking at different images than I am. With the shots I provided there is no way that anyone could argue that LL has less ringing and less artifacts than no LL.
Yes, it seems to me we're looking at different images, too. See here:

http://madVR.com/doom9/stuff/FineSharpComparison1.png
http://madVR.com/doom9/stuff/FineSharpComparison2.png

This is so obvious it's staring me right in the face. How could you not see that? I find these ringing artifacts more obvious that those you're complaining about. So from what I can see, both LL on and off add different kinds of ringing artifacts.

I am talking both about A) as well as B), like I already explained. Not only does the finesharp LL add very heavy ringing artefacts (see the shots above) to the image, but it also darkens the image (on the squirrel image above, see the long hair at the left side, close to the image border, it will get darker).
Darkening some lines/features is not the same as darkening the average brightness of the whole image, which is what I understand as B). And aufkrawall and I were talking about B). And once again, quoting 6233638 and cyberbeing in this context made no sense because they never complained about B), as far as I remember.

-------

Here's a test build using a BT.709 gamma curve with a gamme value of 1.7, instead of a 2.22 pure power curve:

[url]http://madshi.net/madVR8812fs1.rar

Maybe that's an acceptable compromise between LL on and off?

[James Freeman]

Feedback

Super-XBR:
The lines are thicker than NNEDI3, and less sharp.
There is more ringing than on NNEDI3.
The overall image looks a bit smeared.
NNEDI3 wins in all aspects with this image.
http://hookedonhouses.net/wp-content/uploads/2013/05/Blue-House-Lake-Michigan-HB-525.jpg

Finesharp:
A) LL adding ringing artifacts.
B) LL changing the brightness of the original source.

A:
Yes, in tight contrasty spaces, LL adds ringing and artifacts.

B:
LL actually looks much closer to the original overall brightness if the Strength value is high.
Without LL the image is way too bright.
With the test image I posted and strength of 4.0 it is very obvious.

Further, when I use LL with Catmul-Rom when downscaling I see the same haloing artifacts around very contrasty detail as "A" states, but as "B" states, the overall brightness of the image is more accurate.

A and B (behavior of LL) can interfere with each other if the image is too small with lots of contrasty fine lines (like grass).

To conclude, I would prefer LL with weaker strength that no LL.

[madshi]

Super-XBR:
The lines are thicker than NNEDI3, and less sharp.
There is more ringing than on NNEDI3.
The overall image looks a bit smeared.
NNEDI3 wins in all aspects with this image.
Agreed. The original image does already have ringing artifacts in it, once again, though. Which "helps" NNEDI3 win the comparison. NNEDI3 would have won, in any case, but with a smaller lead.

A) LL adding ringing artifacts.
B) LL changing the brightness of the original source.

A:
Yes, in tight contrasty spaces, LL adds ringing and artifacts.

B:
LL actually looks much closer to the original overall brightness if the Strength value is high.
Without LL the image is way too bright.
With the test image I posted and strength of 4.0 it is very obvious.

Further, when I use LL with Catmul-Rom when downscaling I see the same haloing artifacts around very contrasty detail as "A" states, but as "B" states, the overall brightness of the image is more accurate.
Agreed.

-------

Sorry, I think I made a mistake when uploading the test build in my previous comment. Here's a corrected version:

http://madshi.net/madVR8812fs2.rar

[Ver Greeneyes]

http://madvr.com/doom9/stuff/FineSharpComparison2.pngThis is a pretty interesting comparison. So it looks like LL adds more ringing, but the ringing is more symmetric and of lower intensity than no-LL, in a way that preserves the original brightness. It reminds me somewhat of images of different phase response settings for the SoX resampler, e.g. (linear versus intermediate):

http://src.infinitewave.ca/images/Impulse/SoX14_VHQ_LP.pnghttp://src.infinitewave.ca/images/Impulse/SoX14_VHQ_IP.png

I'm kind of curious how FS in a perceptually uniform color space like CAM02-UCS would compare, but such conversions are pretty slow (and dependent on lighting conditions).

[huhn]

Yes, it seems to me we're looking at different images, too. See here:

http://madVR.com/doom9/stuff/FineSharpComparison1.png
http://madVR.com/doom9/stuff/FineSharpComparison2.png

This is so obvious it's staring me right in the face. How could you not see that? I find these ringing artifacts more obvious that those you're complaining about. So from what I can see, both LL on and off add different kinds of ringing artifacts.

in the left part both finesharp LL on and LL off show different problems while i give a huge edge to LL on on this part but no finesharp is still better both have ringing problems on this part.

but on the right side LL is clearly bad on this picture and this is a real picture unlike the left part.

[aufkrawall]

Ok, thanks. Is it possible you were testing with v0.88.11 originally and didn't have repair set to 1.0? That might explain why you saw aliasing problems with lower thinning values. I think I need to set repair even higher than 1.0 when using thinning values as high as 0.05. That might get rid of the aliasing.

Repair was set to 1.0. Either it took it over from the new version or I set it manually.


Thinning can be very useful, but it depends on the source and whether you apply it before or after upscaling etc.

Ok, thanks.


In the long run there is not going to be an option, but that doesn't mean I've already fully decided yet which way to set the linear light option. I might force it off instead of on, or try middle ground. Let's wait and see what the discussion will bring.

I think I can confirm iSunrise's worries with this font example:
With LL, types get some "knobs" (aka ringing?) which shouldn't be there.
LL:
http://abload.de/thumb/llumrws.png (http://abload.de/image.php?img=llumrws.png)

GC:
http://abload.de/thumb/gce9ocg.png (http://abload.de/image.php?img=gce9ocg.png)

Fonts look clearly better to me without LL.
Scaling was NNEDI3 64 doubling + Jinc3AR. FineSharp strength was 1 and thinning 0, repair 1, mode 3.

[iSunrise]

@madshi:
I see. I think we are getting closer. Can you please tell us, how you did your examples? Because when I look at them, they are looking almost like pixel art. Are these natively upscaled within madVR and which upscaler did you use?

Just FYI, my shots were done with NNEDI16 doubling on both luma as well as chroma and Jinc for upscaling.

For some reason your example shots look a lot more pronounced than mine, yours are way too sharp, already, so why would you sharpen them even more? Also, when I look at your second example and I enhance it, you can clearly see the thick artefact that appears out of nowhere in the black area at the middle. Personally, this is bothering me way more than anything else, because it seems to come solely from the LL setting, while the LL off setting does exactly show black as black (0,0,0) and not a grey thick bar. This is even more visible, the more you upscale. LL off does not care how much you upscale, there will be no thick artefacts inside.

See:
http://abload.de/img/image5ip2l.jpg

So what do we have?

LL does show a little less ringing on the outside, but instead we do get a thick black bar that appears out of nowhere, irregardless of line art or real-world images. It also acts like a "black hole" like I showed you in the game recording example I did show in my last post. In that example, instead of grey squares that should have sharp, defined edges, we get interrupted edges with black ringing that overlays the borders. This is not acceptable, IMHO.

My point-of-view is that a sharpener should first and foremost help us sharpen the image as a whole (an uniform sharpener, if that's our goal here), therefore one expected negative is that it will produce a bit more ringing on the outside to make it look more sharp, it achieves that exactly by the things you've just shown. However, it should NOT do what finesharp LL does to the image, add a thick artefact, change the original look and also suppress image detail. And that's exactly what I see with your images when LL is ON.

[James Freeman]

I have a nice method to agree on the three finesharp strength values.

First I have a native 1080 (vertical) image, then I'll downscale it to 720 and 480 with lanczos3 in my image viewer, now I have 3 images in different size.
Then I open the native 1080 image in my image viewer (1:1), next to MPC-HC with the smaller image upscaled via madVR laczos3.
Then I enlarge the mpc-hc+madVR image till it matches the size of the original and play with Finesharp till they look "alike".

What you see here in the image I posted is without finesharp, only upscaling lanczos3.

The goal is to find a strength level that will be good for 480 content for "High". and 720 for "Medium".
"Low" will slightly enhance HD and UHD content for sharper look without too much effect.

From my tests, I'd give 3.0 for High, 1.5 for Mid and 0.7 for Low.

Sorry for the huge pic:
http://www.mediafire.com/convkey/984c/a4azi5u5jkwei4lzg.jpg

[ryrynz]

My 750Ti when in FSE mode using the FSE seek bar and jump forward a reasonable margin I get a lot of frame drops and glitches.
This only occurs with D3D 11 in FSE, changing the CPU and GPU queues seems to be the only thing that helps.

[Werewolfy]

Feedback

I've changed my mind. I think we'll make bigger steps faster, if we concentrate the feedback on very specific things. So while you're welcome to do further tests with SuperRes etc, if you like, please concentrate your efforts on FineSharp. I would like to remove all the FineSharp controls, and just end up with low/medium/high (for both "image enhancement" and "upscaling refinement"). When testing FineSharp, it would make sense to disable LumaSharpen and SuperRes, so that you really only test FineSharp separately. You can test FineSharp either in image enhancements (before upscaling) or in upscaling refinement (after upscaling). Testing it before upscaling should have a stronger effect, so it might be easier to see the difference between various settings there. But you decide whether you want to test it before or after upscaling.

Questions:

1) Do you prefer linear light on or off?
2) In my own very short tests I found that FineSharp sometimes introduces aliasing artifacts. These seem to be mostly fixed by setting the "repair" option to rather high values. Personally, I've tried setting "repair" to 1.0, and liked the result. But what is your opinion about this? Do you find "repair" at 1.0 works for you? Or would you prefer it at a lower value?
3) Do you see a difference worth noting between the 3 different modes? Please note that these modes will make more of a difference if the sources have stronger grain. So in order to judge which modes work best and which worst, it might make sense to also test with a source with a lot of grain in it. FWIW, mode 3 is slower, modes 1 and 2 are faster. So if you like mode 3 best, but not much better than 1 and 2, then it would still be useful to know whether you prefer 1 over 2 or the other way round.
4) Which combinations of strength and thinning would you suggest for low/medium/high presets?

Thanks! :)

1) Difficult to say... LL is closer of the original picture but sometimes it darkens the contours especially on animes... I will try your latest test build.
2) I never witness a case where "repair" at 1.0 harms the picture.
4) For image enhancement :
Low : Strength : 0.3. 0.4 works for me too, the effect is just a little bit stronger but I wouldn't want someting higher than 0.4 for a low preset.
Medium : Strength : 0.8
High : Strength : 1.8

I find hard to see a difference with the thinning parameter. I will play with it later and will test Fine Sharp in Upscaling Refinement but I don't really like it for that purpose... Fine Sharp tends to sharpens also the compressions artifacts and the ringing. A deblocker could be useful. Lumasharpen is worse because it amplifies also the noise... A denoiser will help a lot.

My preference goes to Super Res but I will talk about that in details later. Just a quick note : I've read that anti-ringing at 1.0 doesn't harm the picture. It's FALSE! Anything higher than 0.25 creates weird artifacts.

[iSunrise]

LL is closer of the original picture but sometimes it darkens the contours especially on animes
If you have the time, can you please provide three shots with finesharp off, finesharp LL and finesharp with no LL, so that we can see what you just described? It will help madshi and us to make better decisions based on real examples. Especially anime should suffer a lot with LL on, since it will cause heavy ringing artefacts on uniform areas, which animes and cartoons do mostly consist of. The more you need to upscale, the worse it will look.

I agree with your strength settings, I find that 0.3-0.5 (low), 0.6-0.8 (medium) and 0.9-1.2 (high) would suffice a lot of people. Optimally, an ultra high setting with about 2.0-2.5 would also satisfy everyone that wants really high sharpening. Just IMHO, of course.

[Werewolfy]

I just took the 3 pictures that madshi gave us when he talked about that matter. You can clearly see that even if you use very low strength, without LL the picture gets brighter. Even with a strength at 0.1, the effect is there.

But I agree with you, LL is bad for anime.

[Braum]

I like super-xbr a lot, it's very efficient compared to NNEDI3, even if it's not as good concerning IQ.

I got a 7870XT and an i5 3570K@4.5ghz and super-xbr is a great "spot". My rig can't handle NNEDI3 but the load is pretty low with NEDI.

[aufkrawall]

When you criticize FineSharp non LL for highering overall brightness, I think you should compare it to Lumasharpen.
Lumasharpen is still muuch worse in this matter, at least in cartoons.

[madshi]

This is a pretty interesting comparison. So it looks like LL adds more ringing, but the ringing is more symmetric and of lower intensity than no-LL, in a way that preserves the original brightness. It reminds me somewhat of images of different phase response settings for the SoX resampler, e.g. (linear versus intermediate)
Yep, but it all seems to depend on the exact image content. Sometimes LL on has more ringing, sometimes LL off.

in the left part both finesharp LL on and LL off show different problems while i give a huge edge to LL on on this part but no finesharp is still better both have ringing problems on this part.

but on the right side LL is clearly bad on this picture and this is a real picture unlike the left part.
Yes, but things like the left side can happen with a real picture, too.

I think I can confirm iSunrise's worries with this font example:
With LL, types get some "knobs" (aka ringing?) which shouldn't be there.
LL:
[url=http://abload.de/image.php?img=llumrws.png]
GC:
[url=http://abload.de/image.php?img=gce9ocg.png]
Fonts look clearly better to me without LL.
Scaling was NNEDI3 64 doubling + Jinc3AR. FineSharp strength was 1 and thinning 0, repair 1, mode 3.
FYMI: What happens if you invert (white -> black, black -> white) the image? Does that change anything?

I see. I think we are getting closer. Can you please tell us, how you did your examples? Because when I look at them, they are looking almost like pixel art. Are these natively upscaled within madVR and which upscaler did you use?
Very simple setup: Viewed at 200%, using NNEDI3-16 for image doubling. Then FineSharp enabled in upscaling refinement with default settings (2.0, 0.019).

For some reason your example shots look a lot more pronounced than mine, yours are way too sharp, already, so why would you sharpen them even more?
I've simply used the FineSharp default settings. They may be too high, but what does it matter? Using too high settings can help figuring out problems, because artifacts are more pronounced than they would otherwise be. As you can see in my screenshots, LL on and LL off have different kinds of ringing. In some parts LL on is better, in other parts LL off is better. Which is what I said right from the beginning.

Have you tried my test build yet?

LL does show a little less ringing on the outside, but instead we do get a thick black bar that appears out of nowhere
"Little" and "thick" is somewhat subjective, and also depends on the gamma curve of the monitor.

No comment from you about the "diag" image I posted? :p

I have a nice method to agree on the three finesharp strength values.

First I have a native 1080 (vertical) image, then I'll downscale it to 720 and 480 with lanczos3 in my image viewer, now I have 3 images in different size.
Then I open the native 1080 image in my image viewer (1:1), next to MPC-HC with the smaller image upscaled via madVR laczos3.
Then I enlarge the mpc-hc+madVR image till it matches the size of the original and play with Finesharp till they look "alike".

What you see here in the image I posted is without finesharp, only upscaling lanczos3.

The goal is to find a strength level that will be good for 480 content for "High". and 720 for "Medium".
"Low" will slightly enhance HD and UHD content for sharper look without too much effect.

From my tests, I'd give 3.0 for High, 1.5 for Mid and 0.7 for Low.
In theory not a bad idea, if you wanted to manually choose different FineSharp settings for different upscaling factors. However, my plan was that the upscaling refinement settings should be agnostic of the upscaling factor. Meaning that they should adjust themselves, depending on the upscaling factor.

I think the best test would be to do what you did, but using a fixed exact 200% upscaling factor. And then with this fixed upscaling factor pick low/medium/high presets for upscaling refinement, where "high" would try to preserve the original image sharpness. Everything else, like handling different upscaling factors would then be handled automatically by madVR. At least that was my plan...

If you're talking about image enhancements, these settings are supposed to be used for soft sources, regardless of whether or how much you like upscale them. So I would rather try to find some soft DVD or Blu-Ray sources and look for good low/medium/high settings for 1:1 unscaled playback of those soft sources.

My 750Ti when in FSE mode using the FSE seek bar and jump forward a reasonable margin I get a lot of frame drops and glitches.
This only occurs with D3D 11 in FSE, changing the CPU and GPU queues seems to be the only thing that helps.
Do these frame drops and glitches go away automatically after a couple of seconds of playback? Or do they pile up endlessly?

1) Difficult to say... LL is closer of the original picture but sometimes it darkens the contours especially on animes... I will try your latest test build.
2) I never witness a case where "repair" at 1.0 harms the picture.
4) For image enhancement :
Low : Strength : 0.3. 0.4 works for me too, the effect is just a little bit stronger but I wouldn't want someting higher than 0.4 for a low preset.
Medium : Strength : 0.8
High : Strength : 1.8

I find hard to see a difference with the thinning parameter. I will play with it later and will test Fine Sharp in Upscaling Refinement but I don't really like it for that purpose... Fine Sharp tends to sharpens also the compressions artifacts and the ringing. A deblocker could be useful. Lumasharpen is worse because it amplifies also the noise... A denoiser will help a lot.

My preference goes to Super Res but I will talk about that in details later. Just a quick note : I've read that anti-ringing at 1.0 doesn't harm the picture. It's FALSE! Anything higher than 0.25 creates weird artifacts.
Thanks for the feedback. Yes, I've also seen artifacts from SuperRes AR. Personally, I do like to use FineSharp for upscaling refinement, provided you use a good quality source and a good image doubling algorithm. E.g. try a very sharp downscaled image (preferably one which doesn't have ringing artifacts), quadruple resolution to 400%, then enable FineSharp in upscaling refinement. Try turning sharpness down to 0 and play with just the thinning. Then do it the other way round. That should give you a good feeling for what either of these parameters do.

I just took the 3 pictures that madshi gave us when he talked about that matter. You can clearly see that even if you use very low strength, without LL the picture gets brighter. Even with a strength at 0.1, the effect is there.

But I agree with you, LL is bad for anime.
How about the test build?

I like super-xbr a lot, it's very efficient compared to NNEDI3, even if it's not as good concerning IQ.
Yep!

When you criticize FineSharp non LL for highering overall brightness, I think you should compare it to Lumasharpen.
Lumasharpen is still muuch worse in this matter, at least in cartoons.
LumaSharpen is built for speed above all else. It should not be used as a reference of any sort. Increasing overall brightness can not be good, there's no way to argue about that. But added ringing artifacts are not good, either. Probably noticeable ringing artifacts are worse than a mild image brightness change, but as I've proven, LL on in some situations has less ringing artifacts than LL off, so there are many things to be looked at and evaluated to come to an objective conclusion.

Tried the test build yet? :)

[XMonarchY]

Any ideas why Direct3D 11 causes so many problems on my rig? I think in previous version of madVR, it was finally working fine, but in the latest, it keeps creating tiny stutters 10-15 times per 45 minute video. I am sorry if I missed a reply to this...

[ryrynz]

Do these frame drops and glitches go away automatically after a couple of seconds of playback? Or do they pile up endlessly?


On a 1080 clip it does this for about 10-15 seconds with standard queue settings, on a smaller video it only lasted a second or so. Adjusting the GPU /CPU setting seemed give the most benefit. Going beyond 24 (max GPU) didn't really help. Changing the delay playback settings didn't change anything.

Just passing on that on the 750Ti I have to disable the present a frame for every vsync option for D3D 11 or I get a small amount of glitches and dropped frames every so often.

[iSunrise]

Have you tried my test build yet?
Will have to wait until the weekend, I don't have access to my setup currently (iPhone). Will that reduce the heavy artefacts or what exactly is your goal with it? Can you show a before-after? If you can achieve the LL off look inside the square (a perfectly black surface) and leave the brightness outside intact like with LL on, I am all for it.

"Little" and "thick" is somewhat subjective, and also depends on the gamma curve of the monitor.
It's thick relative to the what's left of the black area that originally was a perfectly 0,0,0 black area. The outside ringing on that image is barely noticable in comparison, secondly, the ringing on the outside was already part of the image that just got more pronounced.

No comment from you about the "diag" image I posted? :p
The diag image definitely shows more pronounced outside ringing, I never disagreed with you on that. In fact the above should make that clear. I very rarely watch black (0,0,0) fonts on white backgrounds (255,255,255), though. Not sure about yourself. All the other examples (including the squirrel itself) are still more accurate without LL. Everyone is free to provide more real-world examples. I am pretty sure there's a clear pattern though, since more and more people now even agree that animes can show the negative effects it has. And if they zoom into them or just do higher upscaling, it will get even worse.

[James Freeman]

I have a nice method to agree on the three finesharp strength values.
However, my plan was that the upscaling refinement settings should be agnostic of the upscaling factor. Meaning that they should adjust themselves, depending on the upscaling factor.

I think the best test would be to do what you did, but using a fixed exact 200% upscaling factor. And then with this fixed upscaling factor pick low/medium/high presets for upscaling refinement, where "high" would try to preserve the original image sharpness. Everything else, like handling different upscaling factors would then be handled automatically by madVR. At least that was my plan...


Oh, that's a great idea. Automation is always good!
So for 480 and 720 just scale them 200% and try to match appearance to 1080 right?
Even if MPC-HC will downscale the image to the screen size? OR retain the video frame on "Double Size"?

[aufkrawall]

FYMI: What happens if you invert (white -> black, black -> white) the image? Does that change anything?

Doesn't look good for LL either, fonts look angular and not smooth anymore.

GC:
http://abload.de/thumb/gcw6peg.png (http://abload.de/image.php?img=gcw6peg.png)

LL:
http://abload.de/thumb/ll13pe7.png (http://abload.de/image.php?img=ll13pe7.png)

Edit: Oh, and there is massive ringing with LL.

Gonna try the testbuild.

[madshi]

Any ideas why Direct3D 11 causes so many problems on my rig? I think in previous version of madVR, it was finally working fine, but in the latest, it keeps creating tiny stutters 10-15 times per 45 minute video. I am sorry if I missed a reply to this...
You're not providing enough information for me to say anything useful. Do any of the queues empty when that stuttering happens? Which queue(s)? Do you get frame drops or presentation glitches in that moment?

On a 1080 clip it does this for about 10-15 seconds with standard queue settings, on a smaller video it only lasted a second or so. Adjusting the GPU /CPU setting seemed give the most benefit. Going beyond 24 (max GPU) didn't really help. Changing the delay playback settings didn't change anything.

Just passing on that on the 750Ti I have to disable the present a frame for every vsync option for D3D 11 or I get a small amount of glitches and dropped frames every so often.
Oh, so disabling the "present a frame for every vsync" fixes this issue for you? What's your refresh rate and your movie frame rate? That option should only have an effect when the refresh rate / movie frame rate don't match exactly.

Will have to wait until the weekend, I don't have access to my setup currently (iPhone). Will that reduce the heavy artefacts or what exactly is your goal with it?
As I wrote earlier, I've switched from a 2.2 pure power curve to an 1.7 BT.709 curve. The change in gamma value alone means the test build is somewhere in the middle between LL on and off. The switch to the BT.709 curve further seems to reduce those "black ringing" artifacts. I hope the test build will make a nice compromise everybody can live with.

It's thick relative to the what's left of the black area that originally was a perfectly 0,0,0 black area. The outside ringing on that image is barely noticable in comparison, secondly, the ringing on the outside was already part of the image that just got more pronounced.
Once again "barely noticable" is your subjective impression. To my eyes it's very obvious. And no, the outside ringing around the squirrel "box" was *not* already part of the image. There is zero ringing outside of the black box in the original image. Check my screenshot with FS off again. It's simply incorrect what you're claiming here.

The diag image definitely shows more pronounced outside ringing, I never disagreed with you on that.
Oh yes, you did. Some posts ago I wrote:

> But in the left half of the squirrel test pattern, actually there
> were less ringing artifacts with LL turned on compared to off.

Then you replied with:

> If you refer strictly to the diagonal lines test, please zoom
> into your own images and show me that there is actually less
> ringing, because I could not reproduce that. LL always produces
> darker edges with more heavy ringing.

("diag" shows the same effect as the diagonal lines)

I very rarely watch black (0,0,0) fonts on white backgrounds (255,255,255), though. Not sure about yourself.
Once again simply incorrect information from your side. We're talking about black fonts *and* lines on *gray* background, not white background. I don't have the time atm to go hunting for real image content which has black lines on gray background, but it shouldn't be that rare...

Oh, that's a great idea. Automation is always good!
So for 480 and 720 just scale them 200% and try to match appearance to 1080 right?
Even if MPC-HC will downscale the image to the screen size? OR retain the video frame on "Double Size"?
Well, I think you really need to decide first if you want to test for image enhancements or for upscaling refinement. Either is fine with me, but needs different testing methods. If you want to test upscaling refinement, I'd suggest to double resolution exactly, by setting MPC-HC to "Double Size", and then pick any sharp 480 or 720 source, double resolution and try to find FineSharp settings which more or less restore the original sharpness. That would then be the ideal setting for "high" for upscaling refinement. You may want to play with both "strength" and "thinning", because both are mostly independent of each other and have different effects.

[aufkrawall]

1.7 BT.709 curve looks nice to me:
http://abload.de/thumb/newe6qh9.png (http://abload.de/image.php?img=newe6qh9.png)
http://abload.de/thumb/new0gp5w.png (http://abload.de/image.php?img=new0gp5w.png)

[madshi]

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

Can anybody else besides iSunrise, aufkrawall and me go back to v0.88.11 and compare FineSharp with LL on/off with Anime content (and maybe double check with the test build)? There were several votes in favor of "LL on", so I don't want to revert that decision without some more "support". Thanks.

[aufkrawall]

Thanks for openly listening to feedback. :)

[madshi]

What good would it be to ask for feedback - and then ignore it when it comes? :) Just trying to achieve the best possible image quality. But I always like to look at the good and bad sides of all options first, before making a decision.

[Della]

Thanks for openly listening to feedback. :)

And *always* responding politely.
A class act.

[James Freeman]

Unlike camera captured material, anime is digitally or hand drawn with ideal and maximum sharpness.
Anime is not blurred (low pass filter or MTF) by the lens or electronics to prevent aliasing or moire that is captured by camera.
Anime or cartoons ARE high contrast material with sharp lines filled with color.

IMO, sharpness algorithms are there so that we can compensate for "less than ideal" camera traits.
This is why super-sampling then dowscaling looks sharper, or capturing with 5K camera and downscaling to HD look WAY better than capturing with HD camera to HD (the Hobbit for example).
Sharpness algorithms are there to emulate that.

That is why I think anime or cartoons are no indicator or the right way to tweak a sharpness algorithm which WILL be applied on camera captured movies also.
In other words, Anime, Fonts and Test Patterns are NOT good test images for sharpness algorithms.

[iSunrise]

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

Can anybody else besides iSunrise, aufkrawall and me go back to v0.88.11 and compare FineSharp with LL on/off with Anime content (and maybe double check with the test build)? There were several votes in favor of "LL on", so I don't want to revert that decision without some more "support". Thanks.
Thanks a lot for listening.

If anyone is against it, please provide some samples. I simply objected to make madVR better for all of us and I hope that people agree that is what we all want.

[aufkrawall]

Dam, it's so cool how 1080p game captures look with high bitrate @WQHD, thanks to FineSharp, super-xbr and SuperRes. I'd almost think it is rendered in realtime in native resolution. :eek:
And 60fps aren't any problem performance wise.

[Shiandow]

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

Can anybody else besides iSunrise, aufkrawall and me go back to v0.88.11 and compare FineSharp with LL on/off with Anime content (and maybe double check with the test build)? There were several votes in favor of "LL on", so I don't want to revert that decision without some more "support". Thanks.

I've been following the discussion a bit and the problem might be that you're trying to do everything in linear light. For a simple unsharp mask filter you tend to get better results if you blur the image in linear light but subtract the unsharp mask in gamma light (or maybe even L*a*b). Unfortunately I couldn't immediately figure out how to apply this to finesharp, but at first glance the RemoveGrain11, FineSharpB parts should be done in linear light, for RemoveGrain4 it doesn't matter, so the LL artefacts are probably caused by either FineSharpA or FineSharpC.