Closed dyphire closed 1 year ago
I agree, this is too much blur for this scene. I just pushed a commit (bda61d72b3dc898ba13d07038ebff289c4c74051) that adds a light-noise tuned profile here, let me know what you think. I'll also look into building an slightly lighter profile.
nlmeans_light.glsl
output:
I agree, this is too much blur for this scene. I just pushed a commit (bda61d7) that adds a light-noise tuned profile here, let me know what you think. I'll also look into building an slightly lighter profile.
nlmeans_light.glsl
output:
The effect of nlmeans_light.glsl
has improved in my test, although significant texture loss can still be seen
I pushed a commit that fixes a bug which was creating extra blur. Here is the new state of nlmeans_light.glsl
:
Focusing only on the blinds in the background, it now appears that the difference between -vf=nlmeans
and nlmeans_light.glsl
is quite small.
There is low frequency noise on the left head and the pink suit that is handled well by -vf=nlmeans
but not handled well by any of my nlmeans shaders. So I built an nlmeans profile to replicate the effect of -vf=nlmeans
(it can be found here) and the result is virtually identical to -vf=nlmeans
(I measured an SSIM of 0.997064
where 1.0
would mean identical):
So everything appears to be working correctly now. I'll eventually make light variants for LQ/HQ/sharpen/sharpen+denoise/etc., and the new ffmpeg_eqv
profile will always be there too.
I thought about it some more and I decided to move nlmeans_ffmpeg_eqv.glsl
to the dev/experimental
subdirectory. I updated the link in my previous comment to point to the new location.
When using the shader of nlmeans to test its noise reduction effect in some videos, I noticed that it produces severe texture smearing behavior, which did not occur in its filter version. test file: sample.zip
nlmeans.glsl
vf_nlmeans