madVR - high quality video renderer (GPU assisted)

[TheShadowRunner]

v0.88.11 fixed the crash when returning to windowed mode from fullscreen (on XP), thank you madshi.

[Anima123]

For LR sources NNEDI3 for doubling and NEDI for the second round (with or without SuperRes) produces good quality.

[XRyche]

Originally Posted by Anima123

For LR sources NNEDI3 for doubling and NEDI for the second round (with or without SuperRes) produces good quality.

I actually use NNEDI3 32 neurons for doubling and NNEDI3 32
neurons for quadrupling for 240-360p already. I use NNEDI3 32 for doubling and NEDI for quadrupling for 480p. I tried SuperRes in conjunction with these settings.

[MistahBonzai]

As XRyche noted SuperRes introduces some strange distortions to the image edges (in my case the text overlay) as I cranked up the passes. It appeared to add a somewhat opaque overlay while adding extra edges outside the original text edges (not ringing - just sorta smudged out with faint new edges).

I discovered that SuperRes worked quite well for Chroma upscaling even with bicubic. I had been using LumaSharpen (1.4.1 (sharp) and 1.5.0 shaders with MPC-HC as well as FineSharp within AviSynth but never looked that closely at the results other than sharpening. I also found that LumaSharpen does indeed introduce halos as sharpness is increased. I now try FineSharp first before deciding if I want more detail enhancement.

[Anima123]

Originally Posted by XRyche

I actually use NNEDI3 32 neurons for doubling and NNEDI3 32
neurons for quadrupling for 240-360p already. I use NNEDI3 32 for doubling and NEDI for quadrupling for 480p. I tried SuperRes in conjunction with these settings.

SuperRes does not work as well with NNEDI3 as with NEDI, that may be one cause for your observation.

[JarrettH]

If using superres, does the upscaling algorithm matter as much?

I'm guessing "yes" since jinc will be more faithful to the original image than bicubic or bilinear, then superres will have a better upscaled image to work with.

Edit: Just saw madshi is looking for answers to this question :P

[JarrettH]

cyberscott and huhn seem to be landing on some good settings

0.40 to 0.45 for strength

0.25 sharpness

anti-aliasing (reducing to 0.20 to 0.25 to compensate a bit)

anti-ringing 0.8 to 0.9 (1.0 overcompensates)

in my case it is mainly 720p upscaled to 1080p

I also prefer to "refine image once" since I won't be doing multiple 2x upscales

[XRyche]

Originally Posted by Anima123

SuperRes does not work as well with NNEDI3 as with NEDI, that may be one cause for your observation.

I'll try that and see if it makes a difference. Thanks for the suggestion Anima123.



I just tried it with my test video and it does lessen the effect to some extent but the weird artifacts are still present. For super lo-res videos NNEDI3 32 neurons with no extra sharpening seems to look the best. Maybe if a general denoise/deblock shader or equivalent were run before SuperRes that might elevate the weird artifacts. As it stands now, SuperRes doesn't work for me with super lo-res material. It works fine with standard definition material as long as the material isn't too flawed.

[baii]

Originally Posted by madshi

I don't know. Some more details would be helpful. Which video and target resolutions does it work with and which not? For each such combination please post the "movie resolution" and "target rectangle" information from the debug OSD (Ctrl+J). Just the text, no screenshots, please.

Seem like only works when scale factor >2 when rule is "if (scalingFactor.x > 1 )"

480p to windowed 15,0,2404,1264 or full screen 1,0,2561,1440- works
720p to windowed "" - not works
720p to full screen 1,0,2561,1440 - works
1080p or 800p scale to any size(meaning less than 1440p) - not works

confirmed by toggling options on the fly and pull up setting menu in windowed mode.

edit: hmm.. rule change to "if (scalingFactor.x > 0.9 )" and everything seem to work correct~
edit2: hmm nvm, "if (scalingFactor.x > 0.9 )" isnt it~

[Anime Viewer]

Originally Posted by madshi

feedback

-------

Now that we've summed up debanding, I'd love to hear your feedback about "image enhancements" and "upscaling refinement". That's a rather broad topic, covering FineSharp, LumaSharpen und SuperRes, both before and after scaling. Tons of options, so a lot to test, a lot to play with. Let me know what you think, what you like and what you don't like.

-------

Let me describe again the concept I had in mind for these algorithms: In my experience there are 2 different causes to get soft/blurry video playback:

A) Either the source is already blurry.

If this is the case, a good idea might be to sharpen the source before upscaling it any further. The reason for this is that sharpening algorithms work by analyzing the direct neighborhood of each pixel in the image. Upscaling causes the image to get bigger, which means the neighbor pixels in the original image are much further apart from each other in the upscaled image. But the sharpening algorithm only looks at the near neighborhood. Which means that if we sharpen after upscaling, although the sharpening still somewhat works, it loses much of its effectiveness. If the source is already soft to begin with, sharpening after upscaling won't do very much, so for soft sources my idea was to sharpen them before upscaling.

This is what the "madVR\processing\image enhancement" settings page is for. You can think of "softness" as being just another artifact of some sources, and we fix it by sharpening the source before we scale it. This only applies to soft sources, though, so my suggestion would be to not enable these settings by default, but only by demand, if you have a really soft source.

I like the image enhancing before up-scaling option. It is less taxing and lower render times than applying the same effect (both Finesharp and LumaSharpen) as upscaling refinement options. With my Optimus system I can do SuperRes if I use the Nivida gpu, but not if I'm using the Intel gpu. Finesharp (mode 2) prescaling with image enhancement works for both gpu, and thus is more universal. If I use the Finesharp in image enhancement along with the SuperRes in the upscaling refinement section it seems that it makes the image too sharp for my tastes (adding artifacts/aliasing), but are more acceptable if using them together in one section or the other not mixed between the two sections.
I see similar image quality from Finesharp in image enhancement as to what I get with SuperRes in the upscaling refinement. The Finesharp in image enhancement bumps up the present number slightly and there is no increase in render times while anything in the upscaling refinement leads to increased render times.
Since the image enhancement selections is less taxing and works more universally (regardless of gpu) I'd vote for it over upscaling refinement options.

Originally Posted by madshi

My hope here is that we can find a way to setup the "upscaling refinement" in such a way that every source maintains its natural sharpness after upscaling, regardless of which upscaling factor is used. I would like to find upscaling refinement settings which we can setup once and then forget, settings which work for any source and any upscaling factor.

At least that's the idea. I think the option "refine the image after every ~2x upscaling step" should be the right one to use to realize the above approach. But for testing purposes I want to let you choose to only apply upscaling refinement once after all upscaling is done, so you can compare the effect this option has.

Since the Finesharp (mode 2) used as image enhancement doesn't tax the system it seems like the most forget it and set it choice that will work regardless of video resolution and gpu.
The 2x setting with SuperRes seems to lead to high render settings (18-19ms in 480 video) compared to the refine only once option (13-14ms in 480 video), and I don't know that I saw a visual difference between the two as a result I'd stick with the faster rendering option.

Originally Posted by madshi

Some specific questions I'd like to get an answer for:

1) In my tests medium "error upscaling quality" was quite a bit faster than high, while producing almost the same quality. Do you agree? Can I drop the "high" setting and simply always use "medium"?

I agree I don't see a significant difference that would warrant the high over medium setting, and would chose the medium setting.

Originally Posted by madshi

2) It seems that running FineSharp and LumaSharpen *before* SuperRes has a much smaller effect than running them afterwards. So my question would be: Which in your opinion produces the better results: Running FineSharp / LumaSharpen before or after SuperRes?

Using FineSharp and LumaSharpen in image enhancement with SuperRes doesn't seem to give as good image quality as using the three together in upscaling refinement. If the three (or at least FineSharp and SuperRes) are going to be used together then it seems like they should be used both in upscaling enhancement instead of mixing the two by using FineSharp in image enhancement along with SuperRes in upscaling refinement. If there were a way to run SuperRes in image enhancement (pre-upscale) that like the other image enhancement options it could be less taxing, and work better with the other image enhancement options however given the nature of SuperRes I don't think such an option is plausible.

Originally Posted by madshi

3) Which combination of settings do you like most? SuperRes alone? Or in combination with FineSharp and/or LumaSharpen? Or maybe just FineSharp?

I used to think just FineSharp, or just SuperRes alone was a good option that provided sharping, but not to the point of artifacts or over sharping. However experimenting today LumaSharpen has started to grow on me. Going with any of them alone provides a blurrier picture, but less artifacts. Combining two or more gives a sharper image, but more traces of artifacts. It seems like there is no right answer between the two, and its more up to the viewers preference as to which he/she would rather have: sharper images with the chance of noticeable artifacts, or blurrier images with less noticeable artifacts.

Originally Posted by madshi

4) Does using the upscaling refinement options change your preferred upscaling/doubling algorithm? Which did you prefer without upscaling refinement? Which do you prefer with upscaling refinement?

When using upscaling refinement the upscaling algorithm used seems to make far less difference than if no refinement (or image enhancement for that matter) was selected. While I had preferences prior to the upscaling refinement options being added I no longer have them when either image enhancement or upscaling refinement is active.

Of the SuperRes presets I prefer the image quality/performance of the NEDI preset as opposed to the NNEDI3 presets.

Originally Posted by madshi

P.S: Forgot:

8) Do you prefer "linear light" for FineSharp turned on or off?
9) Do you like the LumaSharpen "experimental limiter" turned on or off?
Thanks!!

When using FineSharp as a image enhancement I seem to prefer it with liner light turned on, but when I'm using FineSharp as an upscaling refinement option I like it better with linear light turned off.
I don't think I am seeing a difference between experimental limiter being unchecked or checked.

[MysteryX]

Originally Posted by XRyche

That's weird. I found SuperRes to make quality considerably worse with super lo-res material (240-360p) . The more passes I have it do the worse the effect. Most of my super lo-res material is DVR/TIVO/VHS/Capture card rips from late 90's to mid 2000's. Needless to say they aren't pristine, tbh though, I don't think you will get pristine image quality at those resolution anyways. If there is any type of blocking or aliasing it makes matters worse. When I try to compensate with increasing softness or anti-alaising it creates very weird artifacts. I lack the knowledge to properly describe it but, it looks like I put plastic wrap over my screen. Strange distortions on edges and such. This is leaving the settings at default. It gets much worse when increasing softness or anti-alaising to compensate.

Here's a sample "bad quality" 288p VCD video.

This is with Jinc3+AR upscaling, with Spline4+AR chroma


This is with Bicubic+AR upscaling and 1 pass of SuperRes with default settings, and Spline4+AR chroma


I prefer the 2nd version with Bicubic+SuperRes.

[Ver Greeneyes]

Originally Posted by MysteryX

Here's a sample "bad quality" 288p VCD video.

This is with Jinc3+AR upscaling, with Spline4+AR chroma


This is with Bicubic+AR upscaling and 1 pass of SuperRes with default settings, and Spline4+AR chroma


I prefer the 2nd version with Bicubic+SuperRes.

Man, that frame is screaming for some sort of deblocking. What's weird to me is that the gradients inside the blocks don't seem to meet up at all at the edges - the image almost looks scrambled. Are the smooth gradients inside each block a result of using debanding?

Still, Super Res does make it look slightly less awful.

[XRyche]

Originally Posted by MysteryX

Here's a sample "bad quality" 288p VCD video.

This is with Jinc3+AR upscaling, with Spline4+AR chroma


This is with Bicubic+AR upscaling and 1 pass of SuperRes with default settings, and Spline4+AR chroma


I prefer the 2nd version with Bicubic+SuperRes.

Ok that were really, really, really bad. I personally don't have anything that bad in my lo-res collection. If I did, I got rid of it long ago. I don't think any kind of post processing would make that watchable for me. Even most of my old VHS rips are in better condition than that. That being said, SuperRes does make it look "less horrible". It's so distorted that just about anything would help it.

I've been using an old R.E.M. video as a test video. I'll post some screenshots of the distortions I'm getting a little latter in the day.

[lanzorg]

Why not implementing madvr preset selector with predifined profiles by gpu performance?
A combobox selector with these for example:
- madVR Low
- madVR Medium
- madVR High

[michkrol]

The possibility of implementing presets was discussed some (rather long) time ago and it's not as easy to realise as it sounds.

My understanding of the matter is as follows:
First of all, some (most?) settings are based purely on user preference.
Second, to get most of your GPU, you should make profiles for different source resolutions and frame rates, which can make a really long list.
Third, your performance depends greatly also on target (screen) resolution.
Fourth, to get the most of madVR, your settings should depend on the type of your content (cartoons/anime/live-action/sports/...) and it's quality (old VHS rips/TV/DVD/BluRay/...).
Fifth, you need to take into account that NNEDI (or rather OpenCL) performs horribly on some overall fast GPUs, because of driver's bugs or old architecture.
The list goes on.

To sum up, you can't make presets that work for everyone on every hardware with every content.
To get even close, you'd need lots of work and users would still use their own, custom settings.

That said, there are lots of madVR setting guides (good and bad ones) on the web and the default settings are a good starting point for your own setup.

[baii]

I find that blocks is much more apparent when video is paused. When the video plays, blocks become less of a concern.
I prefer nnedi3 over jinc+refinement for large scale factor.
@mysteryx can you also try compare say nnedi3 32/64 for that frame?

Sent from my 306SH

[MysteryX]

Originally Posted by baii

I find that blocks is much more apparent when video is paused. When the video plays, blocks become less of a concern.
I prefer nnedi3 over jinc+refinement for large scale factor.
@mysteryx can you also try compare say nnedi3 32/64 for that frame?

Sent from my 306SH

NNEDI3 doesn't work on my computer, but I did pre-process and re-encode these videos with an AviSynth script and it looks MUCH better. I'd love to add SuperRes to that AviSynth re-encoding chain but it's not currently possible.

Here's a 288p VCD video
https://mega.co.nz/#!eYIBVT7B!SEPfYY...G5bzkGa7DMEMbQ

In comparison, here's the 720p re-encoded version with Denoise + EEDI3 + NNEDI3 + Sharpen
https://mega.co.nz/#!6EhU2ArI!GyIjNk...iMU6Jl_KCkOTTE

[JackCY]

Before I can post a v0.88.11 is released and Shiandow deband is removed?
It's useful, it's stronger than high settings of the artifact removal and Shiandow works sort of as clear skin, removing as much or as little as you want of the harshness from images.

The High preset does this a little but only a little, which is fine as otherwise it would be too aggressive.
But for SD video which is often a piece of crap quality the Shiandow 1.00/0.01 + grain is pretty good.
Upscaling with SuperRes refinement is good and using cheap upscale algorithms doesn't really change the quality much compared to when SR is disabled.

NNEDI3, sure can be rather performance hungry but nothing a 280x can't handle. Except that I want to keep the GPU quiet and NNEDI3 tends to load it too much, nonetheless NNEDI3_16+SR works well for upscaling SD, even cartoons.
Using NNEDI3 is pretty much pointless now when there is SR, NNEDI3 only seems to make the image better (pixel peeping) when upscale is 2x or more, mostly cartoons and hard lines.
NEDI seems to offer no improvement at all compared to Jinc3+AR.

SR refine done only once after upscaling is fine, settings can be tuned to your desire in strength and profiles.
BUT please make the step of the upscaling refinements 0.1 instead of 0.01 or let us type the numbers in instead of having 2 buttons with 0.01 increments Even holding them takes forever to get from 0.00 to 1.00. The difference with 0.01 steps is non perceptible.

For me, MPCHC + madVR tends to get stuck when closing MPCHC sometimes, doesn't matter what version of MPCHC or madVR. Simply hangs there and has to be killed.

What I would like to see in madVR is a deblocking filter, although I don't need it, and the Shiandow or similar filter that works similar to clear skin, polishing/smoothing the harsh textures of low res when upscaled.

SR is nice even though I hate sharpen and this is practically a sort of smart sharpen. There sure are crazy better SR but then those most likely wouldn't run in real time.

[MysteryX]

Yeah, NEDI doubling... I'm not seeing any advantage over Jinc+AR

[6233638]

Originally Posted by JackCY

Before I can post a v0.88.11 is released and Shiandow deband is removed?
It's useful, it's stronger than high settings of the artifact removal and Shiandow works sort of as clear skin, removing as much or as little as you want of the harshness from images.

The High preset does this a little but only a little, which is fine as otherwise it would be too aggressive.
But for SD video which is often a piece of crap quality the Shiandow 1.00/0.01 + grain is pretty good.

I only just had the opportunity to start doing some testing today (only had the time spare to even watch one film this whole month) and my initial thoughts are similar to yours.

Madshi's debanding seems very good at low levels, but Shiandow's seems to look more natural at high levels of debanding - especially with the "add grain" option enabled. At higher strengths, Madshi's debanding can start to look artificial.
Though Shiandow's debanding may lose more detail (I need to spend more time tweaking to see whether that is actually the case) it looks natural when that happens - with filmed footage at least. I haven't tested any animated content - where Madshi's deband may still fare better.