Welcome to Doom9's Forum, THE in-place to be for everyone interested in DVD conversion. Before you start posting please read the forum rules. By posting to this forum you agree to abide by the rules. |
1st December 2016, 10:19 | #41061 | Link | |
Registered User
Join Date: Aug 2006
Posts: 2,229
|
Quote:
http://forum.doom9.org/showthread.php?p=1788137 I do call it an issue because it does not affect the R9-280X with all other things equal. It seems to only affect newer cards from Nvidia and AMD. If it is driver related then tracking down the cause will allow it to be reported via driver reporting . |
|
1st December 2016, 11:48 | #41063 | Link |
Registered User
Join Date: May 2012
Posts: 447
|
Looking at Neo-XP's zoomed in comparison, there are some spots where SSIM seems to preserve the structure of the upscale better, like a spot on the edge of the woman's face between her eyebrow and eye (on the right side). Other than that it actually seems slightly softer than Bicubic150, but I think the sharpness overall is somewhat more consistent. So I do think this proves that SSIM can make a visible difference, even if the number of pixel differences for a given image are small. Whether the trade off between letting people use all the power of their GPUs and keeping them from wasting power on inconsequential things is worth it is another question.
__________________
Test patterns: Grayscale yuv444p16le perceptually spaced gradient v2.1 (8-bit version), Multicolor yuv444p16le perceptually spaced gradient v2.1 (8-bit version) |
1st December 2016, 12:17 | #41064 | Link | |
Registered User
Join Date: Jan 2016
Posts: 52
|
Quote:
|
|
1st December 2016, 12:17 | #41065 | Link | |
Registered User
Join Date: Mar 2016
Location: Switzerland
Posts: 140
|
Quote:
I will test all the algorithms with a bad source material. It should be easier too see all the qualities and defaults of each one. |
|
1st December 2016, 13:01 | #41066 | Link |
Registered User
Join Date: May 2012
Posts: 447
|
Do you mean around the bottom right corner of the highlight (the window reflection) in her pupil? Yeah that looks a little weird, might be there in the upscale too though.
__________________
Test patterns: Grayscale yuv444p16le perceptually spaced gradient v2.1 (8-bit version), Multicolor yuv444p16le perceptually spaced gradient v2.1 (8-bit version) |
1st December 2016, 15:33 | #41067 | Link | |
*****
Join Date: Feb 2005
Posts: 5,643
|
Quote:
EVR-CP has similar issues after sleep/hibernate, which can be solved by turning on this option: options > playback > output > reinitialize when changing display Relevant code: https://github.com/mpc-hc/mpc-hc/blo...rPresenter.cpp (starts at line 1627) |
|
1st December 2016, 16:19 | #41068 | Link |
*****
Join Date: Feb 2005
Posts: 5,643
|
Hi madshi, since you are now also working on simplifying the settings GUI, I have some ideas regarding the presets/profiles.
1) Get rid of the profile groups. It is just unnecessarily complex. Just let presets cover the whole group in each section. 2) On the main level in the GUI (e.g. scaling algorithms page) show a list of all presets in that section. Have buttons to remove and duplicate a preset. 3) Always have at least one preset. For example called "default". This one can't be removed. It is used as a fallback. 4) Display all settings pages just once. Have a dropdown box somewhere at the top to select the "current" preset. Simpler GUI and one can quickly compare settings of different presets. Have an entry called "new" at the bottom of the list. When selected, this takes the user to the preset creation page of that section. 5) Don't use a profile selection script, but instead give each preset a boolean statement that determines if it should get loaded. This has two benefits. First of all, presets become fully independent, making them easy to share/exchange. So we can start making "Doom9 community approved (TM)" optimized presets in the near future. Secondly, it is much easier to have a large number of presets. Non-programmers are often already struggling with relatively simple profile selection scripts. Given the increasing popularity of 4k and 60fps video, plus the availability of HDR and 3D functionality in madvr, the amount of presents can be significant in case of the true enthusiasts. 6) Add buttons to import/export a preset. Preferably in a (human readable) text-based format (xml/ini) for easy sharing. 7) Add a right-click menu to the boolean statement field of the preset. This is used to quickly insert commonly used variables (srcWidth, deintFps, etc) and could even have submenus for stuff like "(deintFps > 31)". |
1st December 2016, 16:22 | #41069 | Link | |
Registered User
Join Date: Oct 2015
Posts: 99
|
Quote:
I like NGU a lot but not so much for SD -> 1080p. The picture is too "frizzy" or oversharp. I'll need to play with soften edges, I guess. |
|
1st December 2016, 16:34 | #41070 | Link | |
Registered User
Join Date: May 2012
Posts: 447
|
Quote:
__________________
Test patterns: Grayscale yuv444p16le perceptually spaced gradient v2.1 (8-bit version), Multicolor yuv444p16le perceptually spaced gradient v2.1 (8-bit version) |
|
1st December 2016, 16:47 | #41071 | Link |
*****
Join Date: Feb 2005
Posts: 5,643
|
Yes, of course, the order of the list of presets should preferably be customizable. That is easy to implement (with up/down button for moving a selected preset).
But ideally the loading rule of a preset is specific enough to avoid multiple matches. At least when we start creating presets for others. The idea is to eventually have a few presets for each common type of situation (e.g. 720p23 -> 1080p23) with increasing quality/complexity. Plus a fixed set of sample files for testing. Plus a dedicated topic for testing results. So everyone can compare rendertimes and pick the most suitable preset for their hardware. Ultimately a little tool that automatically sets proper presets. Unless madshi beats us with some kind of automated performance based magic. Last edited by clsid; 1st December 2016 at 16:58. |
1st December 2016, 16:57 | #41072 | Link |
Registered Developer
Join Date: Mar 2010
Location: Hamburg/Germany
Posts: 10,344
|
As a programmer, I like the profile selection scripts. I have one single view where I can see which rules determine which profiles are used. Having to check every single profile and only seeing one line at a time would be quite a step back, imho.
As a compromise it could offer an "advanced" and "simple" mode, where in simple mode it just constructs the script from the individual lines, and in advanced mode, let me write it.
__________________
LAV Filters - open source ffmpeg based media splitter and decoders |
1st December 2016, 17:09 | #41073 | Link |
*****
Join Date: Feb 2005
Posts: 5,643
|
Yes, that would be a reasonable compromise.
Perhaps additionally (at a later stage) a preset "test" function could be added? A test button that open a dialog where you can specify the various parameters such as srcwidth/srcheight (textbox), fps (dropdown), filename, HDR (checkbox), etc. That then shows which preset(s) matches the chosen values. |
1st December 2016, 18:01 | #41075 | Link | |
Registered User
Join Date: Dec 2014
Posts: 1,127
|
Quote:
That's for downscaling. Although, the difference is so small I don't really care.
__________________
HOW TO - Set up madVR for Kodi DSPlayer & External Media Players Last edited by Warner306; 1st December 2016 at 18:04. |
|
1st December 2016, 19:42 | #41076 | Link | |
Registered User
Join Date: Mar 2016
Location: Switzerland
Posts: 140
|
Quote:
Mitchell-Netravali / Catmull-Rom / Bicubic150 / Lanczos3 / Jinc Zoom 100% on FHD screen and lossless quality : And the source (saved by MPC) : Once again, Lanczos3 wins for me. You can easily see there that Bicubic150 produce a lot of aliasing and other ugly artifacts. That is why I use Lanczos3 after luma doubling with NGU and why it should be available in madVR's new versions, for upscaling and downscaling. Last edited by Neo-XP; 1st December 2016 at 21:04. |
|
1st December 2016, 20:19 | #41078 | Link |
Registered User
Join Date: Mar 2016
Location: Switzerland
Posts: 140
|
Maybe I wasn't explicit enough. I used a bad source material to highlight the defaults of each algorithm. When used for luma doubling, NGU will produce some artifacts that will be increased by the algorithm that you put after it, for upscaling and downscaling. Jinc is also very good if you want a clean result, but the rendering times make it not relevant for me.
Last edited by Neo-XP; 1st December 2016 at 20:23. |
1st December 2016, 20:50 | #41079 | Link | |
Registered User
Join Date: Feb 2014
Posts: 139
|
Quote:
Quote:
NGU is best for source content 720p and up scaled to either 1080p or 1440p. I don't have 4k and cannot comment. For SD and lower, nnedi3 is still king (with SR off if source is dirty enough), and could be improved with Add Grain if madshi decides to allow the combination. Last edited by har3inger; 1st December 2016 at 20:56. |
|
1st December 2016, 21:04 | #41080 | Link |
Registered User
Join Date: May 2011
Posts: 164
|
Quote:
|
Tags |
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling |
Thread Tools | Search this Thread |
Display Modes | |
|
|