vs-basicvsr

[lansing]

I don't understand, what was the filter supposed to do??

[Selur]

I'm not sure either which is why is simply took a clip and made some screen shots and asked whether someone has experience with it,..from what I have seen so far each of the models has a different effect.
So far I would say:
deblur GoPro: seems to mainly sharpen and 'add details' (seems to causes artifacts on cartoons)
deblur REDS: strop denosing and sharpening (seems to only suited for cartoon content)
denoise: also seems to be useful for really noise vhs sources or cartoons
derain: milder denoise/degrain, unsure where this could be useful, seems to remove less when used on more detailed content,...
-> so no real clue

[poisondeathray]

for hinet: the main purpose is single image deblurring . It's not a temporal algorithm. The type of blur it's most suited to - is mostly camera motion blur / shutter blur.

"derain" is for removing "rain" (it's not "degrain") . ie. short streaks of white, like rain drops. It actually works ok on random dark film scratches (not persistent , long scratches) if you invert the src.

Another semi popular deblurring algorithm is deblurGANv2 (also works best on camera motion blur)

[lansing]

Originally Posted by poisondeathray

for hinet: the main purpose is single image deblurring . It's not a temporal algorithm. The type of blur it's most suited to - is mostly camera motion blur / shutter blur.

"derain" is for removing "rain" (it's not "degrain") . ie. short streaks of white, like rain drops. It actually works ok on random dark film scratches (not persistent , long scratches) if you invert the src.

Another semi popular deblurring algorithm is deblurGANv2 (also works best on camera motion blur)

Ok thanks. So it's targeting high resolution photo images, not videos.

[poisondeathray]

Originally Posted by lansing

Ok thanks. So it's targeting high resolution photo images, not videos.

Not necessarily; the gopro model is trained on gopro video . The resolution isn't "high", though

All the single image algorithms tend to have some temporal inconsistencies in the end result when used on video .

YMMV for all these research projects

[PatchWorKs]

Investigating Tradeoffs in Real-World Video Super-Resolution (aka RealBasicVSR) has just been released:

https://user-images.githubuserconten...95e8e5e0b8.mp4

https://user-images.githubuserconten...279bf41c18.mp4

https://user-images.githubuserconten...d5576cd370.mp4

https://user-images.githubuserconten...fad27cea77.mp4

[kedautinh12]

Wow, interesting results

[Selur]

Doesn't really look better than BasicVSR++ to me atm. but in case someone ports it to Vapoursynth we will see more.

[Gravitator]

At the end, stripes appear under the left arm in Titanic.

[PatchWorKs]

Originally Posted by Selur

Doesn't really look better than BasicVSR++ to me atm. but in case someone ports it to Vapoursynth we will see more.

Some other "real world" comparisons done with Colab Demo:

[Selur]

Here's a nice example what BasicVSR++ can do:


Cu Selur

[aegisofrime]

Can I get a bit of help here? After finally setting up the dependencies (they are a bit convoluted for this plugin honestly), I keep getting the following error:

vpy  [error]: CUDA out of memory. Tried to allocate 1.24 GiB (GPU 0; 8.00 GiB total capacity; 5.08 GiB already allocated; 0 bytes free; 5.21 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation occurred while getting frame 0
x265 [error]: unable to open input file <P:\ISO\MainMovie\ME_021\OME021_temp\OME021.vpy>

My Script is very simple,

clip = core.avs.DGSource(r"D:\ISO\Test.dgi")
#core.max_cache_size=16384
#core.num_threads = 8
import havsfunc as haf
core.max_cache_size=16384
clip = haf.QTGMC(clip, Preset='Placebo', TFF=True, Denoiser="KNLMeansCL")
from vsbasicvsrpp import BasicVSRPP
clip = core.resize.Bicubic(clip=clip, format=vs.RGBS, matrix_in_s="470bg", range_s="limited")
clip = BasicVSRPP(clip,model=1)
clip = core.resize.Bicubic(clip=clip, format=vs.YUV420P12, matrix_s="470bg", range_s="limited")
clip = core.resize.Spline36(clip, 1440,960)
clip.set_output()

Input is just a 720x480 DVD video, while my GPU is a RTX 3060Ti, so it only has got 8GB of RAM. It should still be enough right? How can I tweak the arguments so that this works at all?

[Selur]

adding:

,fp16=True

and/or

,tile_x=360,tile_y=480

to the BasicVSRPP code should help.

Cu Selur

[PatchWorKs]

For those - like me - who don't own powerful GPUs, I just asked kodxana (aka Mediator2011):

https://github.com/kodxana/VapourSynthColab/issues/1

[PatchWorKs]

Hi there, can someone make some (performances, particularry) tests of VapourSynth Single Image Super-Resolution Generative Adversarial Network (GAN) ?
(note: I tried their colab without success)

Thanks in advance !

[Selur]

What do you have in mind?

[PatchWorKs]

Originally Posted by Selur

What do you have in mind?

Oh, well, it would be great to have a colab "service" where upload (old) SD stuff and download (x265-10bit encoded) 1080p...

[Selur]

Ah okay, sounded like wanted some benachmarks. No clue about colab.

[PatchWorKs]

Originally Posted by Selur

Ah okay, sounded like wanted some benachmarks. No clue about colab.

Oh, well, of course it would be great to have some tests of newest algos:

https://github.com/Weepingchestnut/A...per-Resolution

[poisondeathray]

PP-MSVSR featured in PaddleGAN looks like the next improvement over basicvsr++ . I'm trying to get it to work, but no luck so far with modifying config files and paths

https://github.com/PaddlePaddle/Padd..._resolution.md
https://arxiv.org/pdf/2112.02828.pdf