madVR - high quality video renderer (GPU assisted)

[James Freeman]

This scene from the Chess video looks absolutely wrong with anything but 0% luminance reduction.
Reminds me of the Black/Blue Gold/White dress controversy image that was on the internet lately.



I'm still experimenting.

[leeperry]

Originally Posted by madshi

Taste got nothing to do with it. I've proven in the past that if users provide objective screenshot proof, I'm very willing to change my mind. Unfortunately your screenshot comparison didn't convince me. Which again has nothing to do with taste, but everything to do with that your preferred settings produced very visible artifacts.

Yes, you were kind enough to bring back SR's 2x refinement after I managed to show you screenshots that changed your mind and you're right maybe my good ole lighthouse is not up to snuff this time.

Originally Posted by madshi

Huh? Please give me a list of those several people saying they "thoroughly enjoyed SR's AB".

Well I got one here:

Originally Posted by AngelGraves13

SR AB was nice especially for upscaled MPEG-2 DVDs. It really smoothed out the rough edges in the transfers and made SR look smoother.

^ What the man said, it smoothes out noisy edges and makes SR look way smoother. Couldn't have said it better, I love it in 60Hz FRC

Just how pathetic does it get when it takes forum goers to backup your point lol

Originally Posted by madshi

MPDN got nothing to do with MPC.

Screenshots in its own dedicated thread make it look like MPC and I haven't seen an option in PotP to run MPDN as a VR. Anyway MPDN is explained to be a "modern extensible media player" when I'm seeking a standalone VR to run in PotP because the latter provides seamless playback and a D3D GUI. If it's not based on MPC whatsoever it might have deemed worthy to use a more modern GUI.

[Ver Greeneyes]

Originally Posted by James Freeman

This scene from the Chess video looks absolutely wrong with anything but 0% luminance reduction.

I haven't tested any other scenes, but on this one I lean toward 25% luminance reduction and 75% saturation. 0% luminance reduction looks too exaggerated to me (on the bits where there's a visible difference), but anything above 25% looks too reduced (like the light is taking on the color of the fabric rather than being white).

Which other demos are good to test with by the way? I know some of them have bad/missing metadata and such.

[har3inger]

Originally Posted by madshi

Not exactly, but something that approaches that: Create an empty file named "ShowRenderSteps" in the madVR folder. Then make a screenshot of the OSD with the 2 different madVR versions. That should allow us to see if there's any difference in the configuration of the render steps (shader passes).


Can you please make a screenshot of the Ctrl+J OSD with these "worst quality" settings where the render queue doesn't fill in windowed mode?

Very cool. I didn't know this was a feature.

I've tried to do this, but the OSD text size is too large to see whats going on beyond 11 steps that involve chroma upscaling and 2 passes of chroma SR, and 1 deband angle step. The rest of it is off the screen. Increasing window height makes the font larger, ever since the OSD font update.

Render times for either .20 or .23 go up to ~55 and ~58 ms respectively (~35 and ~41 ms w/o) when ShowRenderSteps is on.

Edit: screenshots of what little there is (nothing), taken at exactly the same frame in same video.

http://imgur.com/a/qrT7O

[madshi]

Originally Posted by Uoppi

Hmm, me personally, I use SR1 or SR2 for every scaling profile and no other refinements. So SR might not have kicked in previously?

Might have. You can test disabling it manually and see if doing that affects render times in either 90.20 or 90.23.

There's one more thing I might have changed: In older versions when there was image doubling -> downscaling, IIRC I ran upscaling refinement on the downscaled final res. I might have changed that to run on the doubled image before downscaling. Which of course would also increase render times, but should also improve image quality.

Originally Posted by James Freeman

This scene from the Chess video looks absolutely wrong with anything but 0% luminance reduction.

[...]

I'm still experimenting.

Originally Posted by Ver Greeneyes

I haven't tested any other scenes, but on this one I lean toward 25% luminance reduction and 75% saturation. 0% luminance reduction looks too exaggerated to me (on the bits where there's a visible difference), but anything above 25% looks too reduced (like the light is taking on the color of the fabric rather than being white).

Which other demos are good to test with by the way? I know some of them have bad/missing metadata and such.

Keep the feedback coming! Would be great if we could find one setting that works as a decent compromise for most/all demos.

@Ver Greeneyes, the Sony HDR demo makes for a good test, too. I don't know if this new setting affects all HDR demos or just some. Haven't done too much testing myself yet.

Originally Posted by leeperry

Well I got one here: [...]

Ok, so it's AngelGraves13 and you. But there were more users voicing their opinion against SR AB, myself included.

Anyway, if anyone wants SR AB returned, I'm willing to look at further screenshot comparisons.

Originally Posted by har3inger

Very cool. I didn't know this was a feature.

I've tried to do this, but the OSD text size is too large to see whats going on beyond 11 steps that involve chroma upscaling and 2 passes of chroma SR, and 1 deband angle step. The rest of it is off the screen. Increasing window height makes the font larger, ever since the OSD font update.

Oh yes, I often have the same problem. I guess I should make the font size smaller if this special feature is enabled.

Maybe you can try setting chroma upscaling to Bilinear without any extra stuff? I'm quite sure the rendering time slowdown got nothing to do with chroma upscaling. That way we'd see more useful information in the screenshots.

Alternatively you could also create a debug log with 0.90.20 and 0.90.23. That would also help.

Edit: Too much cut off. Could you create a debug log for me with both versions, please?

[zoyd]

Originally Posted by madshi

Hehe - thanks! So which luminance vs saturation setting do you prefer (see new option in the HDR settings page)? Would be great if you could check with multiple different demos, though, because what looks good in one scene/video might not look so well in another scene/video.

Yes, a definite improvement but the optimal setting is color dependent. For very saturated colors 75%-90% luminance reduction looks good but for more unsaturated colors, especially to retain gradients in whites above 100 nits, 0% luminance reduction is needed, and you probably want to maintain current saturation in these cases. Perhaps there is a simple way to weight the adjustment like a 1/r^2 type distance from fully saturated?

I also noticed that the compression level has a significant effect on the resulting appearance so it's important to set that to your true peak white value. In my case the current options 265 or 400 nits straddle my display peak white of 350 nits.

[CarlosCaco]

Originally Posted by madshi

There was a user who complained about upscaling refinement settings not becoming active when upscaling by a relatively small scaling factor. I've changed that, which will of course slow things down if upscaling refinement suddenly becomes active when it didn't before.

Hey madshi that was me! Thanks a lot! Did you changed this on .23? Because in .22 upscallinf refinements for that small scalinh factor only activated with supersampling , Gone test tonight!!!

[madshi]

Originally Posted by zoyd

Yes, a definite improvement but the optimal setting is color dependent. For very saturated colors 75%-90% luminance reduction looks good but for more unsaturated colors, especially to retain gradients in whites above 100 nits, 0% luminance reduction is needed, and you probably want to maintain current saturation in these cases. Perhaps there is a simple way to weight the adjustment like a 1/r^2 type distance from fully saturated?

I'm not sure I fully understand. You say "for more unsaturated colors [...] 0% luminance reduction is needed" and "you probably want to maintain current saturation in these cases". That sounds like I should neither touch luminance nor saturation for "more unsaturated colors". But if I don't touch either of those, the pixel will stay out of gamut! I think I'm misunderstanding something.

Originally Posted by CarlosCaco

Hey madshi that was me! Thanks a lot! Did you changed this on .23? Because in .22 upscallinf refinements for that small scalinh factor only activated with supersampling , Gone test tonight!!!

No, I didn't change anything in that area from .22 to .23.

[zoyd]

Originally Posted by madshi

I'm not sure I fully understand. You say "for more unsaturated colors [...] 0% luminance reduction is needed" and "you probably want to maintain current saturation in these cases". That sounds like I should neither touch luminance nor saturation for "more unsaturated colors". But if I don't touch either of those, the pixel will stay out of gamut! I think I'm misunderstanding something.

How does the luminance/saturation adjustment interact with "compress highlights"?

What I see when there is an out-of-gamut white (with some gradients e.g. between 100-4000 nits) is that compress highlights and 0% luminance reduction maintains the gradients in the highlights while 100% luminance reduction does not.

[madshi]

Originally Posted by zoyd

How does the luminance/saturation adjustment interact with "compress highlights"?

What I see when there is an out-of-gamut white (with some gradients e.g. between 100-4000 nits) is that compress highlights and 0% luminance reduction maintains the gradients in the highlights while 100% luminance reduction does not.

Oh well. It's all relatively complex. I can't do everything in perfect quality in real time in a pixel shader, so I'm compromising accuracy a little bit (but hopefully not much).

The calculation of the amount of luminance reduction is based on a gamut mapping "estimate" performed in xyY. This estimate will probably be slightly off, but it shouldn't be off much. But consequently also the luminance reduction amount might be slightly off. Maybe this results in the lost gradients at 100% luminance reduction, but I'm not sure.

Do the gradients also get lost with 75% luminance reduction? Which HDR demo file (and runtime in the file) are you testing this with?

[leeperry]

Originally Posted by madshi

Anyway, if anyone wants SR AB returned, I'm willing to look at further screenshot comparisons.

OK fair enough, what would you expect to see from screenshots then? Obviously AB will make the picture softer so I'm not sure what would change your mind, I'm certainly not saying that SR AB is perfect but upscaling is all about compromises and I like the smoothness SR AB provides in motion. Maybe SR strengths going from 1 to 10 instead of 1 to 4 would hit the spot, especially when 4 is unusable.

[Ver Greeneyes]

Tried the camp demo, still think 25% luminance reduction and 75% saturation looks best. On the tents in the dark it's hard to choose between 25/75 and 50/50 though. I felt like 25/75 left the red tent a bit too bright, and 50/50 darkened it a bit too much. 33.3/66.7 might be a good compromise for that particular scene. 100/0 is right out by the way, causes awful banding on the sun at the beginning of the scene.

Edit: Oh, I should mention I tested this using 120 nits for the display luminance to match my calibration, but 120 nits might still be overbrightening things. 180 nits looks less washed out, though it's hard to say what the 'right' brightness should be.

[zoyd]

Originally Posted by madshi

Do the gradients also get lost with 75% luminance reduction? Which HDR demo file (and runtime in the file) are you testing this with?

0% was best for this case and not much difference with 25% but started losing detail by 50%. pm'd about clip.

@Ver Greeneyes, I don't get any banding at any of the settings. That tent scene amber headlight renders best at 100%/0% while the white headlight renders best at 0%/100% on my display. Those renderings also best match the full CalMAN LUT mapping.

[ShiftyFella]

Originally Posted by madshi

* HDR: added option to define how to handle too bright & saturated pixels

did quick test

sony camp:
100% luminance 0% saturation: http://i.imgur.com/xTousOt.jpg
50%: http://i.imgur.com/55ztgDJ.jpg
0% luminance 100% saturation: http://i.imgur.com/YWe8Txv.jpg

chess demo:

100% Luminance 0% saturation: http://i.imgur.com/VwRWAI9.jpg
50%: http://i.imgur.com/v0S7SSw.jpg
0% luminance 100% saturation: http://i.imgur.com/KfpXHHJ.jpg

100% Luminance 0% saturation: http://i.imgur.com/z0mJ5hj.jpg
50%: http://i.imgur.com/OA13aqp.jpg
0% luminance 100% saturation: http://i.imgur.com/59CMxPp.jpg

To my eyes camp looked better with 100% luminance but for chess demo it was exact opposite and only 100% saturation looked like it should. I think even with 'looked best' option it's still not quite ideal and i can see in some places of the image where other options work better but rest of the picture looks wrong with them. I'm happy that we have options for now

[zoyd]

The chess demo is encoded incorrectly- I wouldn't use it for eval.

[Warner306]

Originally Posted by madshi

madVR v0.90.23 released

http://madshi.net/madVR.zip

* custom output levels can now be adjusted in 0.25

What does this option do?

[Warner306]

Originally Posted by madshi

madVR v0.90.23 released

http://madshi.net/madVR.zip

* HDR: added option to define how to handle too bright & saturated pixels

Can you also define this?

[Ver Greeneyes]

Originally Posted by zoyd

@Ver Greeneyes, I don't get any banding at any of the settings. That tent scene amber headlight renders best at 100%/0% while the white headlight renders best at 0%/100% on my display. Those renderings also best match the full CalMAN LUT mapping.

Maybe banding was the wrong word, but the sun looks odd to me with 100% luminance reduction. Here's a comparison:

100% luminance reduction and 0% saturation
75% luminance reduction and 25% saturation
50% luminance reduction and 50% saturation
25% luminance reduction and 75% saturation
0% luminance reduction and 100% saturation

(I normally use a 3DLUT but I turned it off for this)

[ShiftyFella]

Originally Posted by zoyd

The chess demo is encoded incorrectly- I wouldn't use it for eval.

Are you referring to values being higher then defined peak luminance in metadata? Shouldn't madvr compression account for this or are they being clipped. I don't know how this is being handled by madvr, as for lg clip, I think metadata is wrong rather then encode but I could be wrong.

[zoyd]

Originally Posted by ShiftyFella

Are you referring to values being higher then defined peak luminance in metadata? Shouldn't madvr compression account for this or are they being clipped. I don't know how this is being handled by madvr, as for lg clip, I think metadata is wrong rather then encode but I could be wrong.

The encode has values up to 10,000 nits which is clearly wrong. madVR has no idea how to handle this situation since it has no idea what the values should be. It makes no sense to tune an algorithm to faulty input data.