From Avisynth wiki
Jump to: navigation, search
Author Kurt Prünner
Version 1.5.4
Download LeakKernelDeint_1.5.4.zip
Category Deinterlacing
License GNU General Public License
Discussion Doom9 Thread



This filter deinterlaces using a kernel approach. It gives greatly improved vertical resolution in deinterlaced areas compared to simple field discarding.

If you set the threshold to 0, you can get totally artifact free results (due to lack of thresholding) but with much less loss of vertical resolution than simple field discarding. For optimal results, however, set a motion threshold that allows static areas of the picture to be passed through. In this mode, the kernel-based deinterlacing of the moving areas preserves their vertical resolution compared to simple interpolation.

Additionally, this version contains a function named LeakKernelBob() (adapted from Scharfis_brain's script of the same name) that does full framerate deinterlacing, i.e. it turn 50 fields per second into 50 frames per second.


  • AviSynth 2.5.8 or later
  • Supported color formats: RGB32, YUY2, YV12

Syntax and Parameters

LeakKernelBob (clip, int "order", int "threshold", bool "sharp", bool "twoway", bool "map", bool "linked", bool "debug", int "forceCPU")
LeakKernelDeint (clip, int "order", int "threshold", bool "sharp", bool "twoway", bool "map", bool "linked", bool "debug", int "forceCPU")

order (0-1, default none!)
This parameter defines the field order of the clip. It is very important to set this correctly. Use order=0 for bottom field first (bff). Use order=1 for top field first (tff). You must specify order; LeakKernelDeint throws an exception if you omit this parameter.
It is essential to set the field order properly for correct rendering. Because setting it correctly is so important, you are strongly encouraged not to make assumptions about the field order of a clip, but rather to verify the field order using the following procedure.
To determine the field order, make an AviSynth script that serves the raw clip without any processing. If it were an AVI, then just AviSource() would be used. For our examples, we'll use AviSource(). Add a script line to separate the fields using top field first, as follows:
Now serve the script into VirtualDub and find an area with motion. Single step forward through the motion. Note whether the motion progresses always forward as it should, or whether it jumps back and forth as it proceeds. For example, if the field order is wrong, an object moving steadily from left to right would move right, then jump back left a little, then move right again, etc. If the field order is correct, it moves steadily to the right.
If the motion is correct with AssumeTFF().SeparateFields(), then your field order is top field first and you must set order=1. If the motion is incorrect, then your field order is bottom field first and you must set order=0. If you are want to double check things, you can use AssumeBFF.SeparateFields() to check correct operation for bottom field first.
threshold (0-255, default 10)
This parameter defines the "motion" threshold. Moving areas are kernel-deinterlaced while non-moving areas are passed through. Use the map parameter to tweak the threshold parameter so that just the combed areas of the frame are deinterlaced.
sharp (true/false, default false)
This parameter, when set to true, selects a kernel that provides better vertical resolution and performs some sharpening of the video. For less sharpening but also less vertical resolution, set this parameter to false.
twoway (true/false, default false)
This parameter, when set to true, selects a kernel that includes both the previous and the following fields for deinterlacing. When set to false, the kernel includes only the previous field. The latter one-way kernel is faster, crisper, and gives less blending (this last advantage makes the filter perform better on anime). The twoway parameter is included in case users want to achieve the behavior of previous versions.
map (true/false, default false)
This parameter, when set to true, shows the areas that are "moving" as determined by the threshold parameter and which will be kernel-deinterlaced. Use this parameter to assist in tweaking the threshold parameter.
linked (true/false, default false; YV12/YUY2 only)
This parameter, when set to true, links the motion masks of the chroma and luma planes so deinterlacing will take place in the same spots on both planes. Use it if you still get residual artifacts in only partially deinterlaced areas, but most of the chroma artifacts of version 1.4.0 were eliminated by using a combined motion map for both chroma planes, thus the default of false. You probably also don't want to set this to true if your video has severe rainbowing, as it will deinterlace the luma plane in those spots thus degrading video quality more than any (other) chroma artifacts could.
debug (true/false, default false)
This parameter, when set to true, enables debug ouput via the DebugView utility. Currently, it shows the version number of the filter and, if hints are present from Telecide(), whether frame are hinted as progressive or interlaced. If hints are not present, the debug output shows all frames as interlaced. Also, you'll get a hint if your frames have different pitches which will disable the MMX optimizations, even though I'm not sure if this can even happen in AviSynth.
forceCPU (0-5, default 0)
This parameter forces the filter to only use specific optimizations:
  • 0 ... autodetect
  • 1 ... C++ code only
  • 2 ... Use MMX code where available
  • 3 ... Use MMX & Integer SSE code where available
  • 4 ... Use MMX & SSE code where available
  • 5 ... Use MMX & SSE & SSE2 code where available
NB: since the code currently only uses MMX, using a value higher than 2 won't result in a speedup.


    Initial release 
    Fixed: don't read beyond the end of frames
    Changed: Small speedup by reading coefficients from memory 
    Fixed: motion mask was b0rked due to field order being swapped when building it
    Fixed: threshold was off by one
    Fixed: map=true Y-values clamped to 235 
    Changed: renamed the plugin to LeakKernelDeint on neuron2's request
    Fixed: implicit "AssumeFrameBased" when using LeakKernelBob 

    Fixed: LeakKernelBob's cache range was too low, which slowed it down considerably

Back to External Filters

Personal tools