More performant median averaging filter/plugin/script?

[WalkyrieVanguard]

Still using the classic median plugin, but it's rather dog slow in terms of plugins and well behind what's going on in VS. Does anyone know of any modern methods for achieving median of clips that's a bit more performant?

Plugin I'm currently using below:

http://avisynth.nl/index.php/Median

https://web.archive.org/web/20251121124338/http://avisynth.nl/index.php/Median

[wonkey_monkey]

Expr is about 4-5x faster on YUV444P8:

Expr(clip1, clip2, clip3, "x y min x y max z min max")

(not rigorously tested and may not work properly with chroma)

[WalkyrieVanguard]

Seems to work great, performance increase is stellar. Thank you!!

[wonkey_monkey]

A dedicated AVX plugin might get another 50% over that, if anyone wants to write one (or enough people want me to do so).

[DTL]

For SIMD accelerated fast median of 3 (up to AVX2 ?) there are already MedianBlurTemporal by pinterf - https://github.com/pinterf/MedianBlur2
For very close to median of 4 and more up to AVX512 - vsTTempSmooth(pmode=1) - https://github.com/Asd-g/AviSynth-vsTTempSmooth (if input number of clips is even - duplicate one of the clips and it will get a bit more weight in selection. With enough many clips this skew expected to be small enough).

To use with separated clips - Interleave(all) + median + SelectEvery(center).

[Reel.Deel]

I vote wonkey :) - plus the plugin could use HBD, only RGB64 was added and not optimized

[WalkyrieVanguard]

A dedicated AVX plugin might get another 50% over that, if anyone wants to write one (or enough people want me to do so).

I'd certainly be interested (Ideally with AVX1 and 2 support)

[DTL]

When Asd-g refuses to support an even number of frames in vsTTempSmooth pmode=1 I start to do an AveragePlus() filter https://github.com/DTL2020/AveragePlus with a vector of clips input. Some of the average algorithms support any number of input sources - not only odd numbers (as some of the median average implementations).

So to make a pure median with up to AVX2 someone may take the start of the AveragePlus() filter with vector of clips input and combine with the pinterf implementation of MedianBlur2 filter (median compute engine). It is a simple combination because it only replaces frames fetching from a single input clip to frames fetching from a list of input clips and all other complex median engine parts remain unchanged.

There are lots of different averaging modes planned (from single plane single sample level via many planes same sampling position samples values set to larger logical units averabing like intra frame block based and inter frame frame based). Each averaging unit size mode can support many similarity measurement modes (where applicable depending on unit size/type) and averaging algorithms. For example block-sized median may use SAD/SSIM/VIF type of similarity measurement.