madVR - high quality video renderer (GPU assisted)

[madshi]

To be honest, NG1 does seem to like Mononoke style images particularly well (although I have mostly optimized for real life photos). In other types of images waifu2x is usually clearly sharper.

Lower levels of NG1 are not finished yet, so I can't comment on that too much yet, but I can say that NG1 should pack a punch at any performance level I can manage to scale it to.

There should be a first version available by next weekend, regardless of whether I'm "finished" by then or not. I'll try very hard to release something for you to play with, at least.

[Klownicle]

Originally Posted by nevcairiel

Presumably it works with a 900 series card as well, but 4K support is quite limited without full HDMI 2.0, so if thats your goal you should probably upgrade.

It is my understanding that the 9 series is FULL HDMI 2.0. But HDMI 2.0a introduced HDR, and 2.0b introduces another minor modification. So HDMI 2.0 supports 4k at 60hz. Just not officially HDR. Supposedly the 9 series can be upgraded via a Software upgrade but it has been undetermined that nVidia will go that route. You can have Full 4K support with 2.0 is how I've read it.

My question about upgrading is IF I want to send HDR, I'll need a 2.0a card I imagine, BUT the question is how soon will madVR incorporate sending HDR data directly to the TV. If its not for some time, upgrading now is pointless.

[aufkrawall]

Originally Posted by madshi

There should be a first version available by next weekend, regardless of whether I'm "finished" by then or not. I'll try very hard to release something for you to play with, at least.

Thank you.
Who cares about settings that require less performance than NNEDI3 64?

[nevcairiel]

Originally Posted by Klownicle

It is my understanding that the 9 series is FULL HDMI 2.0. But HDMI 2.0a introduced HDR, and 2.0b introduces another minor modification. So HDMI 2.0 supports 4k at 60hz. Just not officially HDR. Supposedly the 9 series can be upgraded via a Software upgrade but it has been undetermined that nVidia will go that route. You can have Full 4K support with 2.0 is how I've read it.

https://developer.nvidia.com/high-dy...ay-development

NVIDIA Support for HDR
All NVIDIA GPUs from the 900 and 1000 series support HDR display output. The presence of HDMI 2.0 provides the bandwidth necessary for the higher quality signal desirable for HDR. The 1000 series adds support for HDR over future Display Port revisions.

[JarrettH]

Originally Posted by madshi

To be honest, NG1 does seem to like Mononoke style images particularly well (although I have mostly optimized for real life photos). In other types of images waifu2x is usually clearly sharper.

Lower levels of NG1 are not finished yet, so I can't comment on that too much yet, but I can say that NG1 should pack a punch at any performance level I can manage to scale it to.

There should be a first version available by next weekend, regardless of whether I'm "finished" by then or not. I'll try very hard to release something for you to play with, at least.

What ever you call it, please make it easily forum searchable

[Sunset1982]

Originally Posted by madshi

I'll try very hard to release something for you to play with, at least.

That would be nice. Can't wait to try it out and give you a little feedback.

[StinDaWg]

Originally Posted by aufkrawall

Who cares about settings that require less performance than NNEDI3 64?

I do. 60 fps videos require low frame times (~16ms before frames drop). I have a ton of 720p60 content.

I absolutely cannot do NNEDI3 64 neurons with 60 fps videos on my mid level gpu at anywhere close to 16ms, so I'm hoping that this NG1 "low" still gives good quality at Jinc/super-xbr type gpu loads.

[Warner306]

I don't want to buy a brand new graphics card just to use one algorithm in madVR. NNEDI3 64 is too slow for a GTX 750 Ti. I would rather live with what that card can do than drop $300 on a GPU. Many madVR users are not gamers.

[huhn]

i don't understand why performance is an issue here.
even if your current GPU is to slow the new algorithm isn't running away.
what is so hard to live without this new algorithm for now?

[aufkrawall]

Originally Posted by StinDaWg

I do. 60 fps videos require low frame times (~16ms before frames drop). I have a ton of 720p60 content.

It was just a joke. madshi said that he's working on lower quality modes which should be performance-equal to Jinc.
Let's just be happy about better quality first.

[Shiandow]

Originally Posted by madshi

For comparison purposes, here are the results with the latest MPDN version:

MPDN NNEDI3 256 taps + SuperRes (strength 1.0, 10 passes) -- | -- MPDN NNEDI3 256 taps + SSSR (sharp, 2 passes)

Also have a look at how the various algos handle the top and bottom letterbox edges.

FWIW the 'hybrid' mode of SSSR gives a somewhat more nuanced result, it still doesn't come close to NG1 though.

[ryrynz]

Originally Posted by Warner306

I don't want to buy a brand new graphics card just to use one algorithm in madVR. NNEDI3 64 is too slow for a GTX 750 Ti. I would rather live with what that card can do than drop $300 on a GPU. Many madVR users are not gamers.

I simply can't understand this "logic". I'm considering updating my GTX 960 to a 1060 for exactly this purpose.. An extra $100 or whatever to run this new scaler comfortably for most content at low power levels? You'd be mad not to.
Or you know, you could just keep using NNEDI3 or whatever, your choice. Just don't QQ about not wanting to spend money because you can't see the benefit when it's clearly obvious.
There's been enough discussion about more demanding algorithms over the years. Do you really think the upgrade train stops here?

Because Warner306 is having a cry, I vote for:

Next Gen Mad Hardware Upgrade Scaler.

Seriously impressive result there Madshi. Does SuperRes work any good with it, probably not really needed though.

Considering how well this is doing with lines thickness I might actually no longer need to use Avisynth upscaling to achieve similar quality for low res content.
Only a few days to go..

[CarlosCaco]

Originally Posted by ryrynz

I simply can't understand this "logic". I'm considering updating my GTX 960 to a 1060 for exactly this purpose.. An extra $100 or whatever to run this new scaler comfortably for most content at low power levels? You'd be mad not to.
Or you know, you could just keep using NNEDI3 or whatever, your choice. Just don't QQ about not wanting to spend money because you can't see the benefit when it's clearly obvious.
There's been enough discussion about more demanding algorithms over the years. Do you really think the upgrade train stops here?

Because Warner306 is having a cry, I vote for:

Next Gen Mad Hardware Upgrade Scaler.

Seriously impressive result there Madshi. Does SuperRes work any good with it, probably not really needed though.

Considering how well this is doing with lines thickness I might actually no longer need to use Avisynth upscaling to achieve similar quality for low res content.
Only a few days to go..

I think madVR goal is to get the most benefit that you can with your hardware...
For sometime I only had laptops with very weak gpus but even with that I can use madVR to get a very good video playback...
Now I have an desktop with an GTX 570 that can do some more things but is quite limited too...
I don t game at all but I consider video playback a passion, a great hobby so in my opinion worth spend some money on a better gpu only for madVR
Here were i live all hardware things are quite expensive comparing to others country so every upgrade that i want to do on my rig takes a lot of time to happen...
But the goal is even with a old hardware you can use madVR to really improve playback
It's a expensive hobby if you want to go deeper
But all human passions have a cost...

But I have in my update list first a decent gpu, second a good colorimeter, third a new tv
It's take time to get all of this... but this don t make my experience with madVR less enjoyable



Enviado do meu iPhone usando Tapatalk

[YxP]

Originally Posted by CarlosCaco

I think madVR goal is to get the most benefit that you can with your hardware...

As a person who is not going to upgrade his gpu just for madvr I have to disagree. Madvr has always been about that +5% or even +1% upgrade to picture quality. The reason the new algorithm is "next generation", is that it might actually be more than just 1% or 5% upgrade. A lot more to my eyes, and I don't even have huge screens like many here do.

[ryrynz]

We'll see how the lower performance NG1 stacks up soon enough. Geforce 960s are going cheap and handle 64 neuron performance cost easily enough for all 24-30 FPS content I've tested.
Pick one up second hand with a decent cooler on it and your away laughing, probably the best card for madVR at their price point.

Consider that people would've spent thousands on hardware to achieve this a decade ago.. Many of us have dedicated HTPCs for this now.

Originally Posted by YxP

As a person who is not going to upgrade his gpu just for madvr I have to disagree. Madvr has always been about that +5% or even +1% upgrade to picture quality. The reason the new algorithm is "next generation", is that it might actually be more than just 1% or 5% upgrade.

If you have something 750Ti+ level performance then you're going to achieve high neuron NNEDI3 + SuperRes type image quality at low NNEDI3 neuron levels. This will be a good bump up in quality for many and will no doubt be the recommended scaler going forward.

Originally Posted by YxP

I don't even have huge screens like many here do.

Huge screens aren't necessary for best results, it does depend on how close you are to it though.

[CarlosCaco]

Originally Posted by YxP

As a person who is not going to upgrade his gpu just for madvr I have to disagree. Madvr has always been about that +5% or even +1% upgrade to picture quality. The reason the new algorithm is "next generation", is that it might actually be more than just 1% or 5% upgrade. A lot more to my eyes, and I don't even have huge screens like many here do.

At moment my old tv is broken so I using a tv that's not even full hd... but with the material that I have I can see clearly much more than 5% improve... for example downscaling with ssim in case of lower resolution tv... but like I said it's a hobby to me... some people just use the standard settings and get a very good quality picture, some people are more sensible to some aspects like smoothness or artifacts... other not, like for example my brother only starts to perceive stutter and judder because I said and show him... until this he aways watched movies with evr... now his starts to use madVR e became very sensible to this things


Enviado do meu iPhone usando Tapatalk

[aufkrawall]

Until now our problem has been that we can't have everything good at the same time (little ringing & aliasing, good sharpness, no bloating/other artifacts).
NG1 seems to fully aim for that gap. And considering that NNEDI3 32/64 + SR3 is considered pretty much the most reasonable compromise regarding coming close to ground truth and being a good option for bad sources at the same time, it should also save us a lot of GPU power.
Imagine the deringing filter would work better with high SR strength. It would be extremely expensive, combined with NNEDI3 64 + SR3 (or even 4). And probably still look worse than NG1, compared to the screenshots.

[nautilus7]

Originally Posted by HillieSan

Why Next Generation? It will be an improvement, not a next generation thing. It would suggest to name it after the creator. If the scaler is aiming at 4K then I would suggest something like madshi4K1 or madScaler4K1 or simpler mad4K1. In a few years we might get 8K with new possibilities and a new madScaler8K1 scaler. Just a thought.

madshi called it NG1, Next Generation 1. You have to pay attention on these things, if you're about to make a comment.

[ryrynz]

I'm curious about how NG1 came about more than anything. It does look to be based on NNEDI3.. did tritical have a hand in this? Hopefully a somewhat in depth post is coming..

In similar news "Google has shared details on a new image enhancement technique it has developed called RAISR. RAISR, or rapid and accurate image super-resolution, incorporates machine learning to produce high-quality versions of low-quality images.

With RAISR, the team employs machine learning to train on pairs of images, one low quality, one high, to find filters that, when applied to selectively to each pixel of the low-res image, will recreate details that are of comparable quality to the original. Essentially, RAISR uses high-quality images as a benchmark from which its algorithm can establish ways to adjust a low-resolution image."

Likely similar to NG1, considering how close the upscaled image is to ground truth.

[aufkrawall]

I think the upscaled Mononoke image would be even closer to ground truth if a sharper downscaler would have been used.