eedi3vk2

EEDI3 works by finding the best non-decreasing (non-crossing) warping between two lines by minimizing a cost functional. The cost is based on neighborhood similarity (favor connecting regions that look similar), the vertical difference created by the interpolated values (favor small differences), the interpolation directions (favor short connections vs long), and the change in interpolation direction from pixel to pixel (favor small changes).

Requirement

A Vulkan 1.4 capable GPU and driver.

Usage

eedi3vk2.EEDI3(vnode clip, int field[, bint dh=False, int[] planes=[0, 1, 2], float alpha=0.2, float beta=0.25, float gamma=20.0, int nrad=2, int mdis=20, bint hp=False, int vcheck=2, float vthresh0=32.0, float vthresh1=64.0, float vthresh2=4.0, vnode sclip=None, vnode mclip=None, int device_index=0, bint list_device=False, int num_streams=2])

• clip: Clip to process. Any format with either 8-16 bit integer or 16/32 bit float 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.

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

• alpha/beta/gamma: These trade off line/edge connection vs artifacts created. alpha and beta must be in the range [0,1], and the sum alpha+beta must be in the range [0,1]. alpha is the weight given to connecting similar neighborhoods. The larger alpha is the more lines/edges should be connected. beta is the weight given to vertical difference created by the interpolation. The larger beta is the less edges/lines will be connected (at 1.0 you get no edge directedness at all). The remaining weight (1.0-alpha-beta) is given to interpolation direction (large directions (away from vertical) cost more). So the more weight you have here the more shorter connections will be favored. Finally, gamma penalizes changes in interpolation direction. The larger gamma is the smoother the interpolation field between two lines (range is [0,inf]). If lines aren't getting connected then increase alpha and maybe decrease beta/gamma. Go the other way if you are getting unwanted artifacts.

• nrad/mdis: nrad sets the radius used for computing neighborhood similarity. Valid range is [0,3]. mdis sets the maximum connection radius. Valid range is [1,40]. If mdis=20, then when interpolating pixel (50,10) (x,y), the farthest connections allowed would be between (30,9)/(70,11) and (70,9)/(30,11). Larger mdis will allow connecting lines of smaller slope, but also increases the chance of artifacts. Larger mdis will be slower. Larger nrad will be slower.

• hp: A speed vs quality option. hp=True, use half pel steps, hp=False, use full pel steps.

• vcheck/vthresh0/vthresh1/vthresh2/sclip:

  vcheck settings:

      0 - no reliability check
      1 - weak reliability check
      2 - med reliability check
      3 - strong reliability check

  If vcheck is greater than 0, then the resulting interpolation is checked for reliability/consistency. Assume
  we interpolated pixel 'fh' below using dir=4 (i.e. averaging pixels bl and cd).

       aa ab ac ad ae af ag ah ai aj ak al am an ao ap
                            eh          el
       ba bb bc bd be bf bg bh bi bj bk bl bm bn bo bp
                fd          fh          fl
       ca cb cc cd ce cf cg ch ci cj ck cl cm cn co cp
                gd          gh
       da db dc dd de df dg dh di dj dk dl dm dn do dp

  When checking pixel 'fh' the following is computed:

        d0 = abs((el+fd)/2 - bh)
        d1 = abs((fl+gd)/2 - ch)

        q2 = abs(bh-fh)+abs(ch-fh)
        q3 = abs(el-bl)+abs(fl-bl)
        q4 = abs(fd-cd)+abs(gd-cd)

        d2 = abs(q2-q3)
        d3 = abs(q2-q4)

        mdiff0 = vcheck == 1 ? min(d0,d1) : vcheck == 2 ? ((d0+d1+1)>>1) : max(d0,d1)
        mdiff1 = vcheck == 1 ? min(d2,d3) : vcheck == 2 ? ((d2+d3+1)>>1) : max(d2,d3)

        a0 = mdiff0/vthresh0;
        a1 = mdiff1/vthresh1;
        a2 = max((vthresh2-abs(dir))/vthresh2,0.0f)

        a = min(max(max(a0,a1),a2),1.0f)

        final_value = (1.0-a)*fh + a*cint


    ** If sclip is supplied, cint is the corresponding value from sclip. If sclip isn't supplied,
       then vertical cubic interpolation is used to create it.

• mclip: A mask to use edge-directed interpolation only on specified pixels. Pixels where the mask is 0 are generated using cubic linear or bicubic interpolation. The main goal of the mask is to save calculations.

• device_index: Index of the Vulkan device to use.

• list_device: If True, raises an error listing the available Vulkan devices and their indices.

• num_streams: Number of streams to run kernels in parallel on the Vulkan device.

Installation

pip install -U vapoursynth-eedi3vk2