vapoursynth-cdef

VapourSynth port of Constrained Directional Enhancement Filter.

Usage

from vscdef import cdef

clip = cdef(clip)

Reference

The Python wrapper vscdef.cdef provides exactly the same features as the VapourSynth plugin core.cdef.CDEF.
However, the VapourSynth plugin requires 12-bit or 10-bit integer input, and requires mod8 input. The Python wrapper converts it for you.

cdef(
    clip:         vs.VideoNode,

    # All these parameters accept 1 integer for all planes, 2 integers for luma and chroma, or 3 integers for Y, U, and V.
    pri_strength: int | Sequence[int] = [3, 1],
    sec_strength: int | Sequence[int] = [1, 0],
    pri_damping:  int | Sequence[int] = 2,
    sec_damping:  int | Sequence[int] = 1,
    # If the default is too strong, try disabling secondary filtering by setting `sec_strength` to `0` first.  
    # If the default is too weak, try increasing `pri_strength` (of the Y plane) first.  

    planes:       PlanesT             = [0, 1, 2]
)

For each parameters, check:

• https://github.com/5fish/SVT-AV1/blob/main/Docs/Appendix-CDEF.md
• https://arxiv.org/pdf/1602.05975

Note that even though the AV1 codec only supports certain value ranges, the implementation itself supports a much wider range of values.

Examples

On a bad source with heavy mosquito noise, this might be useful.
Note that on newer versions of vs-jetpack with MVUtensils, unfortunately the support for SADMode.ADAPTIVE_SATD_DCT is dropped. You can either modify it to do a different search, or rewrite it using raw MVTools calls if your setup doesn't support it.

from vsdenoise import MotionMode, MVTools, Prefilter, SADMode
from vscdef import cdef

dn = cdef(clip)

mv = MVTools(clip, search_clip=Prefilter.DFTTEST)

mv.analyze(tr=2, blksize=32, overlap=16, truemotion=MotionMode.COHERENCE, divide=2)
mv.recalculate(thsad=14, blksize=8, overlap=4, dct=SADMode.ADAPTIVE_SATD_DCT, truemotion=MotionMode.COHERENCE)

dn = mv.degrain(dn, clip, tr=2, thsad=24)