Skip to content


Folders and files

Last commit message
Last commit date

Latest commit



37 Commits

Repository files navigation


NNEDI3 is an intra-field only deinterlacer. It takes in a frame, throws away one field, and then interpolates the missing pixels using only information from the kept field. It has same rate and double rate modes, and works with YV12, YUY2, and RGB24 input. NNEDI3 is also very good for enlarging images by powers of 2.

Ported from AviSynth plugin and borrowed some codes from &


The file nnedi3_weights.bin is required. On Windows, it must be located in the same folder as NNEDI3CL.dll. Everywhere else it can be located either in the same folder as, or in $prefix/share/NNEDI3CL/. The build system installs it at the latter location automatically.

nnedi3cl.NNEDI3CL(clip, int field[, bint dh=False, bint dw=False, int[] planes=[0, 1, 2], int nsize=6, int nns=1, int qual=1, int etype=0, int pscrn=2, int device=-1, bint list_device=False, bint info=False])
  • clip: Clip to process. Any planar format with either integer sample type of 8-16 bit depth or float sample type of 32 bit depth is supported.

  • field: Controls the mode of operation (double vs same rate) and which field is kept.

    • 0 = same rate, keep bottom field
    • 1 = same rate, keep top field
    • 2 = double rate (alternates each frame), starts with bottom
    • 3 = double rate (alternates each frame), starts with top
  • dh: Doubles the height of the input. Each line of the input is copied to every other line of the output and the missing lines are interpolated. If field=0, the input is copied to the odd lines of the output. If field=1, the input is copied to the even lines of the output. field must be set to either 0 or 1 when using dh=True.

  • dw: Doubles the width of the input. It does the same thing as Transpose().nnedi3(dh=True).Transpose() but also avoids unnecessary data copies when you scale both dimensions.

  • planes: Sets which planes will be processed. Planes that are not processed will contain uninitialized memory.

  • nsize: Sets the size of the local neighborhood around each pixel (x_diameter x y_diameter) that is used by the predictor neural network. For image enlargement it is recommended to use 0 or 4. Larger y_diameter settings will result in sharper output. For deinterlacing larger x_diameter settings will allow connecting lines of smaller slope. However, what setting to use really depends on the amount of aliasing (lost information) in the source. If the source was heavily low-pass filtered before interlacing then aliasing will be low and a large x_diameter setting wont be needed, and vice versa.

    • 0 = 8x6
    • 1 = 16x6
    • 2 = 32x6
    • 3 = 48x6
    • 4 = 8x4
    • 5 = 16x4
    • 6 = 32x4
  • nns: Sets the number of neurons in the predictor neural network. 0 is fastest. 4 is slowest, but should give the best quality. This is a quality vs speed option; however, differences are usually small. The difference in speed will become larger as qual is increased.

    • 0 = 16
    • 1 = 32
    • 2 = 64
    • 3 = 128
    • 4 = 256
  • qual: Controls the number of different neural network predictions that are blended together to compute the final output value. Each neural network was trained on a different set of training data. Blending the results of these different networks improves generalization to unseen data. Possible values are 1 or 2. Essentially this is a quality vs speed option. Larger values will result in more processing time, but should give better results. However, the difference is usually pretty small. I would recommend using qual>1 for things like single image enlargement.

  • etype: Controls which set of weights to use in the predictor nn.

    • 0 = weights trained to minimize absolute error
    • 1 = weights trained to minimize squared error
  • pscrn: Controls whether or not the prescreener neural network is used to decide which pixels should be processed by the predictor neural network and which can be handled by simple cubic interpolation. The prescreener is trained to know whether cubic interpolation will be sufficient for a pixel or whether it should be predicted by the predictor nn. The computational complexity of the prescreener nn is much less than that of the predictor nn. Since most pixels can be handled by cubic interpolation, using the prescreener generally results in much faster processing. The prescreener is pretty accurate, so the difference between using it and not using it is almost always unnoticeable. The new prescreener is faster than the old one, and it also causes more pixels to be handled by cubic interpolation.

    • 1 = old prescreener
    • 2 = new prescreener (unavailable with float input)
  • device: Sets target OpenCL device. Use list_device to get the index of the available devices. By default the default device is selected.

  • list_device: Whether to draw the devices list on the frame.

  • info: Whether to draw the OpenCL-related info on the frame.


Requires Boost.

meson build
ninja -C build