video remastering

[10362227]

Remastered Vid + Source DL:https://1drv.ms/u/s!AhqDWZobqnZ6dPs0ZqVHgfygp6k?e=yVnWLw

The total effect is very very amazing. How does he make it? VapourSynth→AI is it possible ??
I tried IVTC→nnedi3→AI to fix it, but the effect is very bad

Compare: https://imgsli.com/MTIwNTQ0
https://imgsli.com/MTIwNTQz

[ChaosKing]

I would start with one of these
https://github.com/Kiyamou/VapourSynth-RealCUGAN-ncnn-Vulkan
https://github.com/Kiyamou/VapourSynth-SRMD-ncnn-Vulkan
https://github.com/Kiyamou/VapourSynth-RealSR-ncnn-Vulkan
https://github.com/HolyWu/vs-realesrgan
https://github.com/HolyWu/vs-basicvsrpp

And then there is also https://github.com/rlaphoenix/VSGAN + https://upscale.wiki/wiki/Model_Database

[10362227]

I would start with one of these
https://github.com/Kiyamou/VapourSynth-RealCUGAN-ncnn-Vulkan
https://github.com/Kiyamou/VapourSynth-SRMD-ncnn-Vulkan
https://github.com/Kiyamou/VapourSynth-RealSR-ncnn-Vulkan
https://github.com/HolyWu/vs-realesrgan
https://github.com/HolyWu/vs-basicvsrpp

And then there is also https://github.com/rlaphoenix/VSGAN + https://upscale.wiki/wiki/Model_Database

I‘v tried most of them, The result is not good.
Compare: https://imgsli.com/MTIwNTQ0 (https://imgsli.com/MTIwNTQ0)
https://imgsli.com/MTIwNTQz

[MysteryX]

Impressive result indeed! Following.

[Selur]

just for the fun of it here's what:
SwinIR (real_sr x4) looks like: https://imgsli.com/MTIwNjY2
RealESRGANx4plus looks like: https://imgsli.com/MTIwNjY3
ReaslSR looks like: https://imgsli.com/MTIwNjY4
DPIR+SwinIR_leightweight_sr looks like: https://imgsli.com/MTIwNjY5 (script: https://pastebin.com/UTN3pkJQ)
KNLMeansCL+SwinIR_leightweight_sr looks like: https://imgsli.com/MTIwNjc5
=> so looking a the ESRGAN based resizers is probably the right direction,...
If on were to train a model specifically for such videos there should be a bit more possible,...

[Jukus]

Typical https://www.topazlabs.com/ ?

In my opinion, at this stage of AI development, the best thing you can do to improve video is this:
IVTC
QTGMC(video, Preset='very slow', Sharpness=1.0, InputType=1)
nnedi3 rpow2
any sharpening filter
add some grain

Although I see that many people like Topaz (or anything else AI), I personally see that it only makes the video worse, just a daub

[10362227]

Typical https://www.topazlabs.com/ ?

In my opinion, at this stage of AI development, the best thing you can do to improve video is this:
IVTC
QTGMC(video, Preset='very slow', Sharpness=1.0, InputType=1)
nnedi3 rpow2
any sharpening filter
add some grain

Although I see that many people like Topaz (or anything else AI), I personally see that it only makes the video worse, just a daub

No. His Remastered videos for sale, the price is expensive. I have tried many ways to remaster, but the results ...

[ChaosKing]

Who would buy slightly less shitty quality? lol

[shph]

I agree with Jukus. Most heavy AI resizers are bad for movies. Some AI work well for cartoons. AI processing is too slow.
Here is my usual workflow example:

Overwrite Input Scan Type: BFF
TIVTC FM speed slowest, QTGMC very slow
Color - Matrix Rec.601 to Rec.709
DeBlock - QED 34 36
DeGrain - SMDegrain TR2 (this temporal denoise filter hides a lot of artifacts and makes things look better)
DeRing - HQDering (optional. in most cases for cartoons only)
DeHalo - EdgeCleaner (optional. in most cases for movies only)
Resizer - NNEDI3 8x4 N128 N2
Sharpen - CAS 0.8 (fine tune depending of output resolution)
AddGrain - GrainFactory3 (optional. this will bring back some fine texture and hide possible banding if use 8-bit codec. But same time grain will make codec compression less effective and increase final file size)

[ChaosKing]

What is TIVTC FM?

[Selur]

TFMs slow-option <> speed of the field matching function
slow -

Sets which field matching fuction is used. Possible settings:

0 - normal (should have the worst accuracy)
1 - slower
2 - slowest (should have the best accuracy)

This is basically a quality vs. speed option (the slower the function the more
accurate it should be). Prior to v0.9.9.1 of tfm this was a boolean value, true
from previous versions corresponds to a current value of 1 and false corresponds
to a current value of 0.

Default: 1 (int)
In Hybrid it's a TIVTC option named 'FM speed'. :)

Cu Selur