51

(7 replies, posted in Using SVP)

Kyuri wrote:

But what do you mean by add .glsl into mpv.conf?

mpv.conf from Anime4k contains scripts that tells it to load the Anime4k .glsl files from shaders folder:
# Optimized shaders for higher-end GPU: Mode A (HQ)
glsl-shaders="~~/shaders/Anime4K_Clamp_Highlights.glsl;~~/shaders/Anime4K_Restore_CNN_VL.glsl;~~/shaders/Anime4K_Upscale_CNN_x2_VL.glsl;~~/shaders/Anime4K_AutoDownscalePre_x2.glsl;~~/shaders/Anime4K_AutoDownscalePre_x4.glsl;~~/shaders/Anime4K_Upscale_CNN_x2_M.glsl"

mpv.conf from SVP fresh installation contains some script that allows mpv to connect to SVP:
https://gyazo.com/a8457c661de44b303c747a8f6e10bb55

so when you install Anime4k by replacing the old mpv.conf you remove the script that allows it to detect SVP.

I've combined the default mpv.conf file with the one mentioned from darkwinscm's link and now it should contain both the script that detects SVP and the script that tells it to load Anime4k shaders.
All you have to do is unzip it in "C:\Program Files (x86)\SVP 4\mpv64\" and should work if you're using Windows:
https://drive.google.com/drive/folders/ … OuWgBfmLxY

lurker wrote:

Launched for 3840x2160@25fps. The image shakes in some scenes.

The tree wiggle issue from here that was only visible on high resolutions on v2 models? (that gets fixed if u upgrade TensorRT)
https://www.svp-team.com/forum/viewtopi … 617#p84617
Or another issue? Please show an example.

According to https://github.com/hzwer/Practical-RIFE … 2268877015
there is a scenario where the image moves incorrectly. Try maybe model 4.7 v2 instead.

flowreen91 wrote:
jdg4dfv7 wrote:

1) For me18.v2 is by far the best model regarding all things mentioned.

I agree, somehow v19 and v20 managed to break something and now they stretch the images instead of smoothly moving them vertically:
https://github.com/hzwer/Practical-RIFE … 2253162650

They fixed it in v21:
https://github.com/AmusementClub/vs-mlr … e_v4.21.7z

54

(10 replies, posted in Using SVP)

Only difference i see compared to my logs is that your TensorRT enables:
[08/05/2024-01:30:54] Preview Features: kFASTER_DYNAMIC_SHAPES_0805 [ON],
and then crashes:
[08/05/2024-01:31:01] [W] [TRT] Using kFASTER_DYNAMIC_SHAPES_0805 preview feature.
[08/05/2024-01:32:17] [E] Error[2]: [deconvolutionV2Builder.cpp::nvinfer1::builder::CaskDeconvolutionBuilderBase<class nvinfer1::builder::CaskDeconvolutionV2Builder,class nvinfer1::rt::task::CaskDeconvolutionV2Runner,-2147483603>::createDeconvolution::759] Error Code 2: Internal Error (Assertion isOpConsistent(deconvolution.get()) failed. Cask deconvolution isConsistent check failed.)
[08/05/2024-01:32:18] [E] Error[2]: [builder.cpp::nvinfer1::builder::Builder::buildSerializedNetwork::751] Error Code 2: Internal Error (Assertion engine != nullptr failed. )

this page mentions that it was removed eventually:
https://github.com/NVIDIA/trt-samples-f … /issues/65

I would suggest to try and update your TensorRT from 8.5.1 to 9.1.0 using these instructions:
https://www.svp-team.com/forum/viewtopi … 674#p83674
No clue if this will fix anything.
Maybe SVP devs have encountered that error before...

55

(10 replies, posted in Using SVP)

lurker wrote:

(copied the folder and replaced the script)

Uuuh i think the script from that link is outdated.
Use the script from here for TensorRT9 instead:
https://drive.google.com/drive/folders/ … SzUp1SAzbx
Press ` in mpv to see if any error text appears.

jdg4dfv7 wrote:

1) For me18.v2 is by far the best model regarding all things mentioned.

I agree, somehow v19 and v20 managed to break something and now they stretch the images instead of smoothly moving them vertically:
https://github.com/hzwer/Practical-RIFE … 2253162650

chesley96 wrote:

Can anyone give any insight to the scene change detection?

There was a weird scenario recently where only "image comparison" algorithm detected a specific scene change:
https://www.svp-team.com/forum/viewtopi … 102#p85102
You can compare the scene changes yourself by pressing the < and > buttons in mpv.

58

(2 replies, posted in Using SVP)

Sach wrote:

RTX 4090

The settings from UI are enough.
But you can play around with other relevant settings if you're really really bored:
https://www.svp-team.com/forum/viewtopi … 352#p83352

Blackfyre wrote:

You can read more about them on their pages. But also, if you want an article explaining them, as well as better shaders that are more demanding which I cannot run with an RTX 3090 while using RIFE, have a read through this: https://artoriuz.github.io/blog/mpv_upscaling.html

I tried to download ArtCNN since it says here that it beats FSRCNNX:
https://artoriuz.github.io/blog/mpv_ups … ng_results

When starting a video on my RTX 4080, it froze my mpv for 120 seconds to create a cache somewhere. Once it finished, it loads instantly every time.
Try to get it from here and wait for it to finish caching ArtCNN_C4F32_CMP for 3-6 minutes and it should probably run ok:
https://www.svp-team.com/forum/viewtopi … 264#p84264

donnieyeen wrote:

Did anybody try to manually update the TRT libraries to v15.1: latest TensorRT library? Any benefits over the standard TRT install from SVP?

According to Chainik, we should not upgrade yet cause we only get "10-15% performance drop":
https://www.svp-team.com/forum/viewtopi … 612#p84612
For now i would recommend to upgrade with TRT 9.2 from here which fixes a pixels shake issue on rife (v2) models (IF you are using those):
https://www.svp-team.com/forum/viewtopi … 795#p84795

dawkinscm wrote:

Nowadays I mostly use in-built scalers.

I found this link with upscaler examples really useful:
https://github.com/LitCastVlog/Plex-GLSL-Shaders

the ultimate goal is to improve visual quality, so we will always try to pick the combination that makes it look better overall
can u please explain which in-built scalers combination u use so we can try that too?
i preffer to set max visual quality + resize down until GPU stops dieing, than watch pure image native 24 fps. xD

Chainik wrote:

flowreen91
> When trying to transcode ... to see the difference between scene changes

Hint: press '.' key in mpv to step frame by frame. Or "Ctrl+right" in MPC-HC.

By spamming '.' in following example we can see that the cape scene frames are correctly repeated only for the Image comparison algorithm:

SVP motion vectors:
https://gyazo.com/e72e6ff3a090edd396057d0b26eee3f7
NVOF motion vectors:
https://gyazo.com/1b5a7afb1ceabf2ecd0f4a9c6b250de8
Image comparison:
https://gyazo.com/050c30df626410bde53291b9d8c47f76

Is there anything we can do to increase NVOF and SVP algorithms strength in order for them to correctly detect it and not interpolate that cape scene change?

Chainik wrote:

it's clearly a scene change and SVP do everything right - see f1.jpg before SC and f2.jpg after SC

When trying to transcode with following settings https://gyazo.com/219cab646d448fdd8b58d99d07504a2d with Repeat Frames to see the difference between scene changes i'm getting the following:

DoctorStrange2.SVP_sc06nvofmotionvectors.mp4 and DoctorStrange2.SVP_sc99nvofmotionvectors.mp4 have same file sizes
meaning changing rife_sc value doesn't increase the scene detection strength and they both can't detect the cape scene to repeat it.

DoctorStrange2.SVP_sc06svpmotionvectors.mp4 and DoctorStrange2.SVP_sc99svpmotionvectors.mp4 have same file sizes
meaning changing rife_sc value doesn't increase the scene detection strength and they both can't detect the cape scene to repeat it.

DoctorStrange2.SVP_sc15imagecomparison.mp4 doesn't detect the cape scene to repeat it.
DoctorStrange2.SVP_sc06imagecomparison.mp4 detects the cape scene and repeats correctly the frame to skip interpolating it.

Questions:
1. is that cape scene not detectable by nvof and svp motion vectors? cause it always tries to interpolate it, creating 10 frames of rife.jpg
2. or since they are new functionalities they aren't correctly passed to transcoding functionality?
3. if works ok, is there a way to increase/decrease scene detection strength for nvof/svp motion detector until it detects the cape scene as a scene change to be able to repeat those frames?

Thanks!

You can check the examples here:
https://drive.google.com/drive/folders/ … tb3OOUgkz7

64

(5 replies, posted in Using SVP)

SamE wrote:

If i try to uninstall and reinstall the whole SVP thing, will i lose all my setting and profiles then?

My local SVP4 settings and profiles are located here:
"C:\Users\username\AppData\Roaming\SVP4\settings"
Try to make a backup of it first.

dawkinscm wrote:

Doesn't setting it to 100% effectively turn it off.

Scene change 100% allows rife to interpolate everything all the time.
Scene change at 99% and below enables the selected scene change detector to compare the frames
if a scene change is detected then it will prevent the interpolation by repeating the same frame and creating a "microstutter" in order to prevent visual issues that happen when camera teleports around.

dawkinscm wrote:

I tested using an MPV build from this week with the latest SVP update.

With the MPV update, your mpv.conf file reseted, probably you should apply your previous mpv optimizations to get back to the old performance (if u had any).

New Rife 4.18 here:
https://github.com/AmusementClub/vs-mlr … nal-models

This is soooo cool!
https://gyazo.com/cd0cc66a7e77fc893d0d3cc4e8298b89

gerappa wrote:

Before the topic started i used avisynth

SVP devs please provide the avisynth rife.dll too with same fix applied. Maybe it won't have the seek crash.
Thanks!

Chainik wrote:

.dll attached

gerappa's reaction:
https://gyazo.com/a7a6cfd662e826dac89d6b08adf9bd4c

Chainik wrote:

why exactly you want this?

We want this:
1. to just keep stuff up to date whenever possible
2. to see if gerappa can use 2 threads if that line was turned into a warning:
more specifically it mentions on this commit that they changed the crashing "gpu_thread must be between 1 and 1 inclusive" into a warning:
https://github.com/styler00dollar/Vapou … r9_mod_v14
3. maybe u can recompile the old SVP compatible vs_rife.dll file with that fix so gerappa can check if it unlocks that extra thread?

gerappa wrote:

Change the DLL also didn't worked at all (actually froze the video completely).
just the stupid driver not allow it/limit it (if i understands well). But why?

@SVP devs, for us to correctly use the latest rife_vs.dll file from https://github.com/styler00dollar/Vapou … n/releases do we need to update something else too?

gerappa wrote:

Is there anything i can do about it?

Refund the Radeon RX 7900 XT and get any cheap RTX.
Or downgrade the driver like gerappa managed to fix it here:
https://www.svp-team.com/forum/viewtopi … 001#p84001

According to this video it says that installing "Driver Only" prevents issues:
https://www.youtube.com/watch?v=xCX3acP3At4&t=128s
Does the GPU thread issue persists even in this case?

gerappa try to go to
https://github.com/styler00dollar/Vapou … n/releases
and download each windows .dll from recent releases, rename it to rife_vs.dll and replace the initial "C:\Program Files (x86)\SVP 4\rife\rife_vs.dll"

more specifically it mentions on this commit that they changed the crashing "gpu_thread must be between 1 and 1 inclusive" into a warning:
https://github.com/styler00dollar/Vapou … r9_mod_v14

not sure if it would actually fix anything, but please take a look.
Thanks!

jdg4dfv7 wrote:

@flowreen91
Made the line change ("True"), nothing works.

Try with these:
https://drive.google.com/drive/folders/ … SzUp1SAzbx

jdg4dfv7 wrote:

So I understand it was the correct way to do?

Sounds like you made it work, well done!

From what i see if i revert the following step, the old SVP vsmlrt.py file still works for all models including latest 4.17_lite.

Step 5: Edit helpers.py replace line 52
old TensorRT 8.5:
return RIFE(clip,multi,1.0,None,None,None,model_num,backend,ensemble,implementation)
new TensorRT 10.1:
return RIFE(clip,multi,1.0,None,None,None,model_num,backend,ensemble,True,implementation)

And i didn't see any signficant improvements if i replace it with latest vsmlrt.py file from last week:
https://github.com/AmusementClub/vs-mlr … /vsmlrt.py

So for now we can continue using the old file until somebody notices something.

RAGEdemon wrote:

With RIFE ~4.4/4.6, you can easily get 2X - this will up 24fps to 48fps

For minimum loss of quality we should not use RIFE and just go with the SVP's Automatic interpolation because it's not a resource hog.