Question, are those screenshots from the preview window (hybrid) or screenies from the final output file ?
I ask, because it is as if the "strenght' of the denoiser (KNLMeans) has been nerfed, or it is only nerfed in preview mode !!
Thats exactly what i have noticed and see...
That wasn't as such in previous builds of hybrid.
Anyway, i run couple test samples to see if the Quality from the output file is much different from that of what i get to see in the preview window you know..
18.07.2026, 17:08 (This post was last modified: 18.07.2026, 17:09 by Selur.)
Those screenshots are from the Vapoursynth Preview.
Atm. I don't know what you see or what you think is running wrong.
For me, it seems to work as it should. (denoising is working, as demonstrated in the screenshots)
=> If you think there is an issue, explain in detail and share proper details (=> read sticky).
Cu Selur
----
Dev versions are in the 'experimental'-folder of my GoogleDrive, which is linked on the download page.
(18.07.2026, 17:08)Selur Wrote: Those screenshots are from the Vapoursynth Preview.
Atm. I don't know what you see or what you think is running wrong.
For me, it seems to work as it should. (denoising is working, as demonstrated in the screenshots)
=> If you think there is an issue, explain in detail and share proper details (=> read sticky).
Cu Selur
I will be as clear as possible about this subject.
It looks like very litle to no changes are applied using KNLmeans strength → 10.00 🤯 , except the filtered image blurs more details but doesn't enhanced it like it used to be.. !?
See pictar 👇🏼
[img][/img]
Or just download the original picture and zoom in to compare.. It's very obvious something wrong is going on .. ¯\_(ツ)_/¯
(18.07.2026, 06:20)Selur Wrote: => if you would care to share a proper bug report [INFO] Infos needed to fix&reproduce bugs,.. at least the debug output would show me what settings you are using.
Cu Selur
Sure, no problemo 👇🏼
recap: i've loaded media file, toggled QTGMC on , enabled EZdenoise and selected KNLmeans .. Beside that only a color and sharpening is used , thats all.
Look i have a better screenshot taken while using KNLmeansCL seperate filter.
The strength is set to 2.0 in knlmeans, but as you can see something is realy off with knlmeans filter in general
It used to be a much .. MUCH.. cleaner output quality !?
Okay, a few things, you are using QTGMC as denoiser not as deinterlacer.
QTGMC is normally meant to do light clean up or residual combing and similar stuff.
Using something like:
especially EZDenoise=10, will kill a lot of fine details.
DetailSharpen is trying to enhance fine details and contrast, but if you use something like EZDenoise=10 you already kill everything that is there.
So my suggestion is either:
a. move DetailSharpen before the denoising
or
b. lower the Denoising
With EZDenoise=10 it is normal that DetailSharpen or FineSharpen won't find many details to enhance.
Sane values should be 0-3 and normally better 0-1.5.
=> To me Hybrid still does seem to behave as it should.
Cu Selur
----
Dev versions are in the 'experimental'-folder of my GoogleDrive, which is linked on the download page.
Side note: What Hybrid does when 'KNLMeans(CL)' is selected:
1. if no gpu accelleration is enabled it will use vsnlm_ispc
2. if gpu and a nvidia card is available, it will use vsnlm_cuda
3. if a gpu and a only non-nvidia card are available, it will use KNLMeansCL
----
Dev versions are in the 'experimental'-folder of my GoogleDrive, which is linked on the download page.
(18.07.2026, 18:10)Selur Wrote: Okay, a few things, you are using QTGMC as denoiser not as deinterlacer.
QTGMC is normally meant to do light clean up or residual combing and similar stuff.
Using something like:
especially EZDenoise=10, will kill a lot of fine details.
DetailSharpen is trying to enhance fine details and contrast, but if you use something like EZDenoise=10 you already kill everything that is there.
So my suggestion is either:
a. move DetailSharpen before the denoising
or
b. lower the Denoising
With EZDenoise=10 it is normal that DetailSharpen or FineSharpen won't find many details to enhance.
Sane values should be 0-3 and normally better 0-1.5.
=> To me Hybrid still does seem to behave as it should.
Cu Selur
Usualy i never use strength 10 for KNLmeans denoiser. I just used that value to magnify the issue, or trying to expose it rather.
True, i never go above 1,5 - 3.0 max. But then again, in that range usuall the knl denoiser from QTGMC is strongly present . But now, it's as if it's doing nothing at all !!???
You know, knlmeans tends to smoothen image while retaining the sharp contours, much different from making something blurry to mask imperfections / noise you know.
Well, that output effect isn't just as clean anymore like it used to be
Of course, it is also possible that the results could be worse or better depending on the type of video source material I am trying to clean up 🤔.
But i've could have sworn i've been using this particular profile for a long time to clean up SD & even HD content, and always with succes ... Untill now ¯\_(ツ)_/¯
(18.07.2026, 18:17)Selur Wrote:
Quote:Look i have a better screenshot taken while using KNLmeansCL seperate filter.
Okay, not helping either,...
But surely, you can visually see not much gets filtered in the image on your right side .
Again, using KNLMeansCL (seperately) i've used strength 2.0 wich is pretty pretty HIGH, and you ought to see a much smoother image usually , practically much less noise !?
(18.07.2026, 18:17)Selur Wrote:
Quote:The strength is set to 2.0 in knlmeans, but as you can see something is realy off with knlmeans filter in general Huh
2 and 10 *qtgmc internal factor are totally different!
You mean Strength factor from KNLmeansCL (seperately) vs KNLmeans through QTGMC , right?
(18.07.2026, 18:17)Selur Wrote:
Quote:It used to be a much .. MUCH.. cleaner output quality !?
when was that?
I can't remember, but i do remember it's quite a few builds back
(18.07.2026, 18:17)Selur Wrote: Cu Selur
Side note: What Hybrid does when 'KNLMeans(CL)' is selected:
1. if no gpu accelleration is enabled it will use vsnlm_ispc
2. if gpu and a nvidia card is available, it will use vsnlm_cuda
3. if a gpu and a only non-nvidia card are available, it will use KNLMeansCL
I have always used an nvidia gpu with hybrid.
Even when knnlmeans was still working correctly back then .
Well, i sill have an old hybrid build from back in the day. i will install that and see what the result is..🤔
Could it be , something has gone wrong during installation?
Is it usuall safe to install a fresh hybrid revision over an existing installation ?
Please don't tell me that I have to uninstall and remove everything every time a new dev is released.
I do use the "torch🔥" addon pack you know 😰!
Atm. quite a few filters get deprecated and removed or removed by newer/faster version (mainly since the Vapoursynth itself is changing a lot,.. R78 is probably available in a few weeks), so a lot of stuff is changing (I try to keep stuff as backward compatible as possible through wrappers&co), so reevaluating settings might not be the worst idea.
=> uploading a new dev version
Cu Selur
----
Dev versions are in the 'experimental'-folder of my GoogleDrive, which is linked on the download page.
(18.07.2026, 18:47)Selur Wrote: Atm. quite a few filters get deprecated and removed or removed by newer/faster version (mainly since the Vapoursynth itself is changing a lot,.. R78 is probably available in a few weeks), so a lot of stuff is changing (I try to keep stuff as backward compatible as possible through wrappers&co), so reevaluating settings might not be the worst idea.
=> uploading a new dev version
Cu Selur
Ah thanks for the update.
Again, i stick to it.. i remember the Output Quality of KNLmeans used to be much MUCCCHH better... / cleaner