Installation:
mpv.io
Wiki:
github.com
Manual:
mpv.io
User Scripts (including opengl shaders):
github.com
High quality video output profile (goes into mpv.conf):
profile=opengl-hq
mpv.io
Installation:
mpv.io
Wiki:
github.com
Manual:
mpv.io
User Scripts (including opengl shaders):
github.com
High quality video output profile (goes into mpv.conf):
profile=opengl-hq
mpv.io
Other urls found in this thread:
github.com
youtube.com
gpuopen.com
gpuopen.com
github.com
github.com
gist.github.com
arxiv.org
arxiv.org
twitter.com
Did you try FSRCNN?
>github.com
Blur fixed! :)
How to check if we need less than 2x upscaling with mpv to use SSimSuperRes instead of FSRCNN?
Wat.
If 2x or more upscaling, use FSRCNN shader.
If less than 2x upscaling, use SSimSuperRes.
With auto-profile.lua script or course. ;)
I dont think you can do that. You need to create profiles based on video resolution.
:(
Its no big deal. Just create one profile for low resolution content and use SSSR with everything else.
I am probably too stupid for that kind of math.
Replace all 1.500000 with 1.000000 in FSRCNN shader, then it will be used only when upscaling factor is 2x+ (1.5x+ by default).
>1.5x+ by default
Whoops, 1.33 by default.
Oh, i forgot you could do that. I cant find nothing like that in my shaders though.
Does it mean that it downscales just after doubling? It's weird!
Redownload it then.
Thats how doublers work.
Ok! Is it part of the produced quality or perfect doubling is better?
Its 1.500000 actually.
No, that number is not a scaling factor
480*2 / 640 != 640/480
you can't train neural networks to upscale to an arbitrary resolution, so you train them to exactly double the resolution
you then double the resolution however many times you want and then downscale it, and since mpv has a very good downscaler (mitchell) the output is almost perfect
>1.000000
It should be 1.100000 (or 1.000001), don't set it exactly to 1.000000.
How can I make mpv play a youtube playlist?
It should work the same way you play single videos, only instead you point mpv to the playlist (youtube.com
Since I swapped to tiling wm, how do I make mpv pause on left mouse button? Asking for a friend
>[15:06:05] hanna - I guess I'll play around with it if I ever want to implement a fast neural network meme shader in mpv
Which one? Isn't RAVU enough for him?
Anyone else use mpv as an image viewer?
After all, an image is a video with 1 frame
haasn evolves to madshi :D
and a video player is an image viewer using double the resources
do people like haasn have regular jobs or does he only work on mpv?
1. see 2. can't browse pics with arrow (or any other) keys = into the thrash it goes
On windows, if i load a 2 GB movie into mpv, it only uses something like 300 MB of memory.
but how does that work?
Some people said he is sysadmin but others said he is still a student. You should ask him on irc.
size of what you're viewing doesn't matter. What matters is how it is encoded, which there isn't a single stat you can look at to estimate how much resources mpv will use to decode that
anyway I think I need to spend a day with a GCN profiler and gpuopen.com
maybe when I have a vega GPU
because ours is the exact use case that benefits from large thread groups
gpuopen.com
N-nothing can stop haasn now!
but for a quick view it's great
pic rel
Can someone give me some good settings for a piece of shit laptop?
Ewa scalers as fast as bilinear when? ;)
anyone here using spline64 as cscale and scale?
The end of the human race maybe?
No, it's so retarded!
Honeyview shits all over it
>an AMD initiative designed to enable developers to create ground-breaking PC games, computer generated imagery and GPU computing applications for great performance and lifelike experiences using no cost and open development tools and software.
Is he trolling amd once again?
whats options does the opengl-hq setting provide??
Not using QuickLook?
github.com
>not Irfanview
explain why it is retarded. WHY?
t. Irfan
see
JPEGView is the true master race. No UI and everything is inside a config file, just like mpv
>6
tnx
But it's a slow piece of shit.
Quickviewer is much better!
github.com
>JPEGView
>JPEG
does it work with pngs?
how the hell is it slow, exactly? It opens everything instantly
Use a conditional hook
Your jpegshit is slower and fucking dead!
I use mpv as my primary image viewer, yes. It's the only renderer I can trust
Quickviewer is fucking good too! And you can use shaders too!
yeah but that plays the frame and it closes 1 sec after
have to add some configs or somethings
mpv --show-profile=opengl-hq
yeah, right now i open the image with mpv and then mouse right click to pause
Lol
libplacebo when?
right after vulkan, fast FSRCNN and VFIAC ;)
yo my man why you always saying
>;)
So I suppose not before 2030? :(
>him
>he
Honeyview is also calls home :^)
post proof otherwise or stfu
>VFIAC
We wait for VFIAC now. I dont even know what the fuck is that.
All these bloated equations...
No! What the heck?! Human eye cant see above NGU Sharp (high) anyway!
Looks like it does the same thing as SVP but with a deep neural net which means it's probably both a lot better but not gonna be real time any time soon unless someone gets really clever about it.
I wonder if binarization techniques can be applied effectively like in:
arxiv.org
>deep neural net which means it's probably both a lot better but not gonna be real time any time soon
Is it possible to use that to make 23.976 Hz -> 30 Hz so it matches 60 FPS displays? Is it possible to use it to 23.976 Hz -> 60 FPS on low resolution sources?
I exclusively use bpgview and convert all my images to bpg.
Other image formats need to die already.
I mean effectively this just replaces the motion vector estimation by convolution kernel estimation, which is easier to train for and easier to work with. Like with motion vectors, there's no need to estimate the convolution kernel on the full res image - you could do it on a downsampled luma version.
The main problem is going to be _applying_ the convolution on the full sized image. For accurate results, it needs to be big enough to cover the window of interest; which in their application means 41x41. Even at 1080p and using compute shader shared memory, 41x41 is way too slow for realtime; so the main challenge is optimizing this final convolution pass.
The main idea I have in mind is trying to get it to shape the convolution kernel into a particular direction, or perhaps apply an integer offset to the convolution kernel. This would allow us to use a significantly smaller kernel while still benefitting from a good coverage area. (Important for “large” motion distances)
Don't fucking do it. You already have a lot of tscale scalers and already rekt madvr in that aspect. What else do you want? Whats you problem? You're program is already very good as it is and also gains more popularity every day. You can ride Vulkan hype for a while now. Take it easy, why you in such a rush? You already have GNU/Linux to you self. No one really cares about those super duper meme techniques. HDR is the next BIG thing. Everyone likes HDR. Just relax, okay?
Go away madshi
>madshi damage control
If they could get it extremely light on motion artifacts and realistically low on performance cost, that would be absolutely glorious.
I would GLADLY sacrifice scaling detail and use bilinear just to get proper, wobble-free 60FPS. Even if it would take the full power of my computer to do so.
yes please
Nope. I don't think anyone wants that. Just a waste of time in my opinion. LOL. I'm not interested at all. smoothmotion is already perfect. You cant top smoothmotion.
Same. As usual, the hardest part is going to be training the weights. Probably requires some shit like MATLAB/Caffe
Can't you stick to doom9 already FFS
@ all
-not just using custom modes utility or smooth motion for perfect playback
-wanting to have some extremely complicated interpolation technique no player in THE WORLD has yet just to piss off youre competition
WOW
>he thinks madVR still “competes” with mpv
cute
Lol did madshi shot down your pm?
How new here, shitskin?
how come a 4k video lags a lot on youtube but when i stream same video with mpv its smooth and flawless i just cant comprehend how mpv can be this good
>how dare people want a feature i don't like. that makes me ANGEREY!!!!!!
Blurring frames together is only useful if they're already high framerate. Otherwise it's ugly at best.
I couldn't possibly care about pissing off competition or anything of that sort. I just want such things wherever possible.
It would obviously be best if it could be made compatible with all decent players.
Youtube relies on your browser for sync. Browsers have a lot of trouble with sync. Pretty much any standalone player would be superior.
But MPV in particular is also incredibly optimized.
@ 62267505
i only read this thread from time to time usually. no need to be rude.
>watching videos on the browser (javascript,...) vs watching videos on a cross-platform media player.
>expecting a web browser to play video well
The real question you should be asking yourself is how the fuck can browser devs be so BAD
why doesnt firefox use libmpv
Because niggers like you are too retarded to switch on hardware rendering.
62267542
As expected from shitskins. Worthless.
Is it going to play anime?
Even if it did, it would still have to sync it to video output along with its own graphics, such as interface changes, animations, page flows, etc.
Just running it in a separate process is guaranteed better performance.
I wish youtube-dl+mpv would become standard.
It was made for anime~