[HELP] What noise is that?

[Doom83]

Hi, dont know what kind of noise is that, I try a lot of filter but dont go away..only ezdenoise up to 8 but are too strong and ruined video...
I'm not a pro in video editing, I work for personal use, so maybe I dont put the right filter and value...

[Selur]

Looks like ringing artifacts which got compressed too strong.
Assuming this is a reencode if your source and the source does not have so strong compression artifacts, de-ringing filters might help.

Would need an uncropped sample to suggest anything specific. (upload it somewhere and share a link here or send it to me via pm and I can look at it after work)

Cu Selur

Ps: using BasicVSR++:

[Doom83]

Thanks for the reply, here are a shot clip unedited, maybe there are other noise

[Selur]

Had a quick go at it, using aWarpSharp with depth set to 32 and planes set to 'chroma only' helps with the chroma halos.
Either SCUNet (requires torch- or vs-mlrt-addon) or DeblockQED_aWarpSharp-16-onlyChroma-TemporalDegrain2-mlClean-FineDehalo help with the noise.
see: here.
Tweaking the settings and probably using a less compressed source (capturing should always be either without reencoding when capturing streams or lossless) might allow preserving more details.

Problem is, the original source was sharpened (which causes halos and ringing) and overly compressed, which turned the artifacts caused by the sharpening into the mess you have.

Cu Selur

Ps.: BasicVSR++ is too aggressive for the whole image.

[Doom83]

Thanks for your attemp, now I also think its a bit oversharpened..the first sample with scunet look good, maybe look a bit painted...maybe in some close-up in the face make another effect...

So that sample can be better?

[Selur]

Using aWarpsharp2 with Depth 32 on chroma only and then SCUNet, I get:

Using SCUNet and then aWarpsharp2 with Depth 32 on chroma only, I get:


You could add some sharpening or even try CodeFormer(on some sources CodeFormer will perform well, on others not; CodeFormer requires the torchAddon) or maybe both.


If content seems too flat after denoising, try adding some grain.


You could even use some machine learning based upscaler:

(side note: Hybrid currently does not allow switching the resizer per scene and when using machine learning based resizers/filters make sure to check different scene, since their effects can vary a lot depending on the scene/source )

There is ton if stuff you can do, but it will require computational power, tweaking to find good value and time.
=> Happy tweaking and finding the best settings for your preferences.

Cu Selur

Ps.: if you got the time and motivation, you can adjust settings per scene.

[Doom83]

That screen look good, expecially the last 3...

But where are scunet? I dont find in hybrid...

My "goal" is try to clean noise, sharp and generally enhance detail...make image most real...

---

Sometime I use for preset

deinterlace - custom
preset - placebo
match preset - placebo
temporal smoothing - 2
source matching - 2x
sharpness - 1

resize - output 1:1 square pixel from 720x480 at 720x540
resize lanczos

degrain - stpresso
denoise - ccd - 4
denoise - knlmeanscl - channels yuv
derainbow - decross

sharpen - cas - 0.750 chroma only

frame - interpolation - rife v4 double fps (59.94)
frame - resize - nnedi3
line - aaf - edge

But its not perfect...
I see lost details, skin pores in face, cheeks...

[Selur]

But where are scunet? I dont find in hybrid... Confused

Like SCUNet requires the torch-addon which will only work on NVIDIA GPUs.
(the torch-addon is on my GoogleDrive, liked on the download page)

preset - placebo
match preset - placebo

total waste in my point of view

Try adding FineDetail as additional sharpener.

Cu Selur

[Selur]

FineDetail => meant FineSharp or DetailSharpen

expecially the last 3...

I used FineSharp and for the last additionally FeMaSR.

[Doom83]

I've downloaded the file...if I had known it before...
So now I make test from begin...

In the filters there are also this?
GitHub - IceClear/StableSR: [IJCV2024] Exploiting Diffusion Prior for Real-World Image Super-Resolution