Good test pattern: https://www.rtings.com/images/test-materials/2017/chroma-444.png (Compress it with any lossy codec first, for example jpeg. You can do it with mpv, only add screenshot-jpeg-source-chroma=no to mpv.conf). Usage: glsl-shader="~~/KrigBilateral.glsl"

KrigBilateral.glsl

// KrigBilateral by Shiandow
//
// This library is free software; you can redistribute it and/or
// modify it under the terms of the GNU Lesser General Public
// License as published by the Free Software Foundation; either
// version 3.0 of the License, or (at your option) any later version.
// 
// This library is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
// Lesser General Public License for more details.
// 
// You should have received a copy of the GNU Lesser General Public
// License along with this library.

//!HOOK CHROMA
//!BIND LUMA
//!BIND HOOKED
//!SAVE LOWRES_Y
//!WIDTH LUMA.w
//!WHEN CHROMA.w LUMA.w <
//!DESC KrigBilateral Downscaling Y pass 1

#define offset      vec2(0)

#define axis        1

#define Kernel(x)   dot(vec3(0.42659, -0.49656, 0.076849), cos(vec3(0, 1, 2) * acos(-1.) * (x + 1.)))

vec4 hook() {
    // Calculate bounds
    float low  = ceil((LUMA_pos - CHROMA_pt) * LUMA_size - offset - 0.5)[axis];
    float high = floor((LUMA_pos + CHROMA_pt) * LUMA_size - offset - 0.5)[axis];

    float W = 0.0;
    vec4 avg = vec4(0);
    vec2 pos = LUMA_pos;

    for (float k = low; k <= high; k++) {
        pos[axis] = LUMA_pt[axis] * (k - offset[axis] + 0.5);
        float rel = (pos[axis] - LUMA_pos[axis])*CHROMA_size[axis];
        float w = Kernel(rel);

        vec4 y = textureGrad(LUMA_raw, pos, vec2(0.0), vec2(0.0)).xxxx * LUMA_mul;
        y.y *= y.y;
        avg += w * y;
        W += w;
    }
    avg /= W;
    avg.y = abs(avg.y - avg.x * avg.x);
    return avg;
}

//!HOOK CHROMA
//!BIND LOWRES_Y
//!BIND HOOKED
//!SAVE LOWRES_Y
//!WHEN CHROMA.w LUMA.w <
//!DESC KrigBilateral Downscaling Y pass 2

#define offset      vec2(0)

#define axis        0

#define Kernel(x)   dot(vec3(0.42659, -0.49656, 0.076849), cos(vec3(0, 1, 2) * acos(-1.) * (x + 1.)))

vec4 hook() {
    // Calculate bounds
    float low  = ceil((LOWRES_Y_pos - CHROMA_pt) * LOWRES_Y_size - offset - 0.5)[axis];
    float high = floor((LOWRES_Y_pos + CHROMA_pt) * LOWRES_Y_size - offset - 0.5)[axis];

    float W = 0.0;
    vec4 avg = vec4(0);
    vec2 pos = LOWRES_Y_pos;

    for (float k = low; k <= high; k++) {
        pos[axis] = LOWRES_Y_pt[axis] * (k - offset[axis] + 0.5);
        float rel = (pos[axis] - LOWRES_Y_pos[axis])*CHROMA_size[axis];
        float w = Kernel(rel);

        vec4 y = textureGrad(LOWRES_Y_raw, pos, vec2(0.0), vec2(0.0)).xxxx * LOWRES_Y_mul;
        y.y *= y.y;
        avg += w * y;
        W += w;
    }
    avg /= W;
    avg.y = abs(avg.y - avg.x * avg.x) + LOWRES_Y_texOff(0).y;
    return avg;
}

//!HOOK CHROMA
//!BIND HOOKED
//!BIND LUMA
//!BIND LOWRES_Y
//!WIDTH LUMA.w
//!HEIGHT LUMA.h
//!WHEN CHROMA.w LUMA.w <
//!OFFSET ALIGN
//!DESC KrigBilateral Upscaling UV

#define sigma_nsq   256.0/(255.0*255.0)
#define N           8
#define sqr(x)      dot(x,x)

#define M(i,j)      Mx[min(i,j)*N + max(i,j) - (min(i,j)*(min(i,j)+1))/2]

#define C(i,j)      (inversesqrt(1.0 + (X[i].y + X[j].y) / Var) * exp(-0.5 * (sqr(X[i].x - X[j].x) / (localVar + X[i].y + X[j].y) + sqr((coords[i] - coords[j]) / radius))) /*+ (X[i].x - y) * (X[j].x - y) / Var*/)  // commented out part works well only on test patterns
#define c(i)        (inversesqrt(1.0 + X[i].y / Var) * exp(-0.5 * (sqr(X[i].x - y) / (localVar + X[i].y) + sqr((coords[i] - offset) / radius))))

#define getnsum(i)  X[i] = vec4(LOWRES_Y_tex(LOWRES_Y_pt*(pos+coords[i]+vec2(0.5))).xy, \
                                CHROMA_tex(CHROMA_pt*(pos+coords[i]+vec2(0.5))).xy); \
                    w = clamp(1.5 - abs(coords[i]), 0.0, 1.0); \
                    total += w.x*w.y*vec4(X[i].x, X[i].x * X[i].x, X[i].y, 1.0);

#define I3(f, n)    f(n) f(n+1) f(n+2)
#define I9(f, n)    I3(f, n) I3(f, n+3) I3(f, n+6)

vec4 hook() {
    vec2 pos = CHROMA_pos * HOOKED_size - vec2(0.5);
    vec2 offset = pos - round(pos);
    pos -= offset;

    vec2 coords[N+1];
    vec4 X[N+1];
    vec2 w;
    vec4 total = vec4(0);

    coords[0] = vec2(-1,-1); coords[1] = vec2(-1, 0); coords[2] = vec2(-1, 1);
    coords[3] = vec2( 0,-1); coords[4] = vec2( 0, 1); coords[5] = vec2( 1,-1);
    coords[6] = vec2( 1, 0); coords[7] = vec2( 1, 1); coords[8] = vec2( 0, 0);

    I9(getnsum, 0)

    total.xyz /= total.w;
    float localVar = abs(total.y - total.x * total.x) + sigma_nsq;
    float Var = localVar + total.z;
    float radius = 1.5;  // mix(1.5, 1.0, sigma_nsq / Var);

    float y = LUMA_texOff(0).x;
    float Mx[(N*(N+1))/2];
    float b[N];
    vec2 interp = X[N].zw;

    b[0] = c(0) - c(N) - C(0,N) + C(N,N); M(0, 0) = C(0,0) - C(0,N) - C(0,N) + C(N,N); M(0, 1) = C(0,1) - C(1,N) - C(0,N) + C(N,N); M(0, 2) = C(0,2) - C(2,N) - C(0,N) + C(N,N); M(0, 3) = C(0,3) - C(3,N) - C(0,N) + C(N,N); M(0, 4) = C(0,4) - C(4,N) - C(0,N) + C(N,N); M(0, 5) = C(0,5) - C(5,N) - C(0,N) + C(N,N); M(0, 6) = C(0,6) - C(6,N) - C(0,N) + C(N,N); M(0, 7) = C(0,7) - C(7,N) - C(0,N) + C(N,N);
    b[1] = c(1) - c(N) - C(1,N) + C(N,N); M(1, 1) = C(1,1) - C(1,N) - C(1,N) + C(N,N); M(1, 2) = C(1,2) - C(2,N) - C(1,N) + C(N,N); M(1, 3) = C(1,3) - C(3,N) - C(1,N) + C(N,N); M(1, 4) = C(1,4) - C(4,N) - C(1,N) + C(N,N); M(1, 5) = C(1,5) - C(5,N) - C(1,N) + C(N,N); M(1, 6) = C(1,6) - C(6,N) - C(1,N) + C(N,N); M(1, 7) = C(1,7) - C(7,N) - C(1,N) + C(N,N);
    b[2] = c(2) - c(N) - C(2,N) + C(N,N); M(2, 2) = C(2,2) - C(2,N) - C(2,N) + C(N,N); M(2, 3) = C(2,3) - C(3,N) - C(2,N) + C(N,N); M(2, 4) = C(2,4) - C(4,N) - C(2,N) + C(N,N); M(2, 5) = C(2,5) - C(5,N) - C(2,N) + C(N,N); M(2, 6) = C(2,6) - C(6,N) - C(2,N) + C(N,N); M(2, 7) = C(2,7) - C(7,N) - C(2,N) + C(N,N);
    b[3] = c(3) - c(N) - C(3,N) + C(N,N); M(3, 3) = C(3,3) - C(3,N) - C(3,N) + C(N,N); M(3, 4) = C(3,4) - C(4,N) - C(3,N) + C(N,N); M(3, 5) = C(3,5) - C(5,N) - C(3,N) + C(N,N); M(3, 6) = C(3,6) - C(6,N) - C(3,N) + C(N,N); M(3, 7) = C(3,7) - C(7,N) - C(3,N) + C(N,N);
    b[4] = c(4) - c(N) - C(4,N) + C(N,N); M(4, 4) = C(4,4) - C(4,N) - C(4,N) + C(N,N); M(4, 5) = C(4,5) - C(5,N) - C(4,N) + C(N,N); M(4, 6) = C(4,6) - C(6,N) - C(4,N) + C(N,N); M(4, 7) = C(4,7) - C(7,N) - C(4,N) + C(N,N);
    b[5] = c(5) - c(N) - C(5,N) + C(N,N); M(5, 5) = C(5,5) - C(5,N) - C(5,N) + C(N,N); M(5, 6) = C(5,6) - C(6,N) - C(5,N) + C(N,N); M(5, 7) = C(5,7) - C(7,N) - C(5,N) + C(N,N);
    b[6] = c(6) - c(N) - C(6,N) + C(N,N); M(6, 6) = C(6,6) - C(6,N) - C(6,N) + C(N,N); M(6, 7) = C(6,7) - C(7,N) - C(6,N) + C(N,N);
    b[7] = c(7) - c(N) - C(7,N) + C(N,N); M(7, 7) = C(7,7) - C(7,N) - C(7,N) + C(N,N);

    b[1] -= b[0] * M(0, 1) / M(0, 0); M(1, 1) -= M(0, 1) * M(0, 1) / M(0, 0); M(1, 2) -= M(0, 2) * M(0, 1) / M(0, 0); M(1, 3) -= M(0, 3) * M(0, 1) / M(0, 0); M(1, 4) -= M(0, 4) * M(0, 1) / M(0, 0); M(1, 5) -= M(0, 5) * M(0, 1) / M(0, 0); M(1, 6) -= M(0, 6) * M(0, 1) / M(0, 0); M(1, 7) -= M(0, 7) * M(0, 1) / M(0, 0);
    b[2] -= b[0] * M(0, 2) / M(0, 0); M(2, 2) -= M(0, 2) * M(0, 2) / M(0, 0); M(2, 3) -= M(0, 3) * M(0, 2) / M(0, 0); M(2, 4) -= M(0, 4) * M(0, 2) / M(0, 0); M(2, 5) -= M(0, 5) * M(0, 2) / M(0, 0); M(2, 6) -= M(0, 6) * M(0, 2) / M(0, 0); M(2, 7) -= M(0, 7) * M(0, 2) / M(0, 0);
    b[3] -= b[0] * M(0, 3) / M(0, 0); M(3, 3) -= M(0, 3) * M(0, 3) / M(0, 0); M(3, 4) -= M(0, 4) * M(0, 3) / M(0, 0); M(3, 5) -= M(0, 5) * M(0, 3) / M(0, 0); M(3, 6) -= M(0, 6) * M(0, 3) / M(0, 0); M(3, 7) -= M(0, 7) * M(0, 3) / M(0, 0);
    b[4] -= b[0] * M(0, 4) / M(0, 0); M(4, 4) -= M(0, 4) * M(0, 4) / M(0, 0); M(4, 5) -= M(0, 5) * M(0, 4) / M(0, 0); M(4, 6) -= M(0, 6) * M(0, 4) / M(0, 0); M(4, 7) -= M(0, 7) * M(0, 4) / M(0, 0);
    b[5] -= b[0] * M(0, 5) / M(0, 0); M(5, 5) -= M(0, 5) * M(0, 5) / M(0, 0); M(5, 6) -= M(0, 6) * M(0, 5) / M(0, 0); M(5, 7) -= M(0, 7) * M(0, 5) / M(0, 0);
    b[6] -= b[0] * M(0, 6) / M(0, 0); M(6, 6) -= M(0, 6) * M(0, 6) / M(0, 0); M(6, 7) -= M(0, 7) * M(0, 6) / M(0, 0);
    b[7] -= b[0] * M(0, 7) / M(0, 0); M(7, 7) -= M(0, 7) * M(0, 7) / M(0, 0);

    b[2] -= b[1] * M(1, 2) / M(1, 1); M(2, 2) -= M(1, 2) * M(1, 2) / M(1, 1); M(2, 3) -= M(1, 3) * M(1, 2) / M(1, 1); M(2, 4) -= M(1, 4) * M(1, 2) / M(1, 1); M(2, 5) -= M(1, 5) * M(1, 2) / M(1, 1); M(2, 6) -= M(1, 6) * M(1, 2) / M(1, 1); M(2, 7) -= M(1, 7) * M(1, 2) / M(1, 1);
    b[3] -= b[1] * M(1, 3) / M(1, 1); M(3, 3) -= M(1, 3) * M(1, 3) / M(1, 1); M(3, 4) -= M(1, 4) * M(1, 3) / M(1, 1); M(3, 5) -= M(1, 5) * M(1, 3) / M(1, 1); M(3, 6) -= M(1, 6) * M(1, 3) / M(1, 1); M(3, 7) -= M(1, 7) * M(1, 3) / M(1, 1);
    b[4] -= b[1] * M(1, 4) / M(1, 1); M(4, 4) -= M(1, 4) * M(1, 4) / M(1, 1); M(4, 5) -= M(1, 5) * M(1, 4) / M(1, 1); M(4, 6) -= M(1, 6) * M(1, 4) / M(1, 1); M(4, 7) -= M(1, 7) * M(1, 4) / M(1, 1);
    b[5] -= b[1] * M(1, 5) / M(1, 1); M(5, 5) -= M(1, 5) * M(1, 5) / M(1, 1); M(5, 6) -= M(1, 6) * M(1, 5) / M(1, 1); M(5, 7) -= M(1, 7) * M(1, 5) / M(1, 1);
    b[6] -= b[1] * M(1, 6) / M(1, 1); M(6, 6) -= M(1, 6) * M(1, 6) / M(1, 1); M(6, 7) -= M(1, 7) * M(1, 6) / M(1, 1);
    b[7] -= b[1] * M(1, 7) / M(1, 1); M(7, 7) -= M(1, 7) * M(1, 7) / M(1, 1);

    b[3] -= b[2] * M(2, 3) / M(2, 2); M(3, 3) -= M(2, 3) * M(2, 3) / M(2, 2); M(3, 4) -= M(2, 4) * M(2, 3) / M(2, 2); M(3, 5) -= M(2, 5) * M(2, 3) / M(2, 2); M(3, 6) -= M(2, 6) * M(2, 3) / M(2, 2); M(3, 7) -= M(2, 7) * M(2, 3) / M(2, 2);
    b[4] -= b[2] * M(2, 4) / M(2, 2); M(4, 4) -= M(2, 4) * M(2, 4) / M(2, 2); M(4, 5) -= M(2, 5) * M(2, 4) / M(2, 2); M(4, 6) -= M(2, 6) * M(2, 4) / M(2, 2); M(4, 7) -= M(2, 7) * M(2, 4) / M(2, 2);
    b[5] -= b[2] * M(2, 5) / M(2, 2); M(5, 5) -= M(2, 5) * M(2, 5) / M(2, 2); M(5, 6) -= M(2, 6) * M(2, 5) / M(2, 2); M(5, 7) -= M(2, 7) * M(2, 5) / M(2, 2);
    b[6] -= b[2] * M(2, 6) / M(2, 2); M(6, 6) -= M(2, 6) * M(2, 6) / M(2, 2); M(6, 7) -= M(2, 7) * M(2, 6) / M(2, 2);
    b[7] -= b[2] * M(2, 7) / M(2, 2); M(7, 7) -= M(2, 7) * M(2, 7) / M(2, 2);

    b[4] -= b[3] * M(3, 4) / M(3, 3); M(4, 4) -= M(3, 4) * M(3, 4) / M(3, 3); M(4, 5) -= M(3, 5) * M(3, 4) / M(3, 3); M(4, 6) -= M(3, 6) * M(3, 4) / M(3, 3); M(4, 7) -= M(3, 7) * M(3, 4) / M(3, 3);
    b[5] -= b[3] * M(3, 5) / M(3, 3); M(5, 5) -= M(3, 5) * M(3, 5) / M(3, 3); M(5, 6) -= M(3, 6) * M(3, 5) / M(3, 3); M(5, 7) -= M(3, 7) * M(3, 5) / M(3, 3);
    b[6] -= b[3] * M(3, 6) / M(3, 3); M(6, 6) -= M(3, 6) * M(3, 6) / M(3, 3); M(6, 7) -= M(3, 7) * M(3, 6) / M(3, 3);
    b[7] -= b[3] * M(3, 7) / M(3, 3); M(7, 7) -= M(3, 7) * M(3, 7) / M(3, 3);

    b[5] -= b[4] * M(4, 5) / M(4, 4); M(5, 5) -= M(4, 5) * M(4, 5) / M(4, 4); M(5, 6) -= M(4, 6) * M(4, 5) / M(4, 4); M(5, 7) -= M(4, 7) * M(4, 5) / M(4, 4);
    b[6] -= b[4] * M(4, 6) / M(4, 4); M(6, 6) -= M(4, 6) * M(4, 6) / M(4, 4); M(6, 7) -= M(4, 7) * M(4, 6) / M(4, 4);
    b[7] -= b[4] * M(4, 7) / M(4, 4); M(7, 7) -= M(4, 7) * M(4, 7) / M(4, 4);

    b[6] -= b[5] * M(5, 6) / M(5, 5); M(6, 6) -= M(5, 6) * M(5, 6) / M(5, 5); M(6, 7) -= M(5, 7) * M(5, 6) / M(5, 5);
    b[7] -= b[5] * M(5, 7) / M(5, 5); M(7, 7) -= M(5, 7) * M(5, 7) / M(5, 5);

    b[7] -= b[6] * M(6, 7) / M(6, 6); M(7, 7) -= M(6, 7) * M(6, 7) / M(6, 6);

    b[7] /= M(7, 7);
    interp += b[7] * (X[7] - X[N]).zw;

    b[6] -= M(6, 7) * b[7]; b[6] /= M(6, 6);
    interp += b[6] * (X[6] - X[N]).zw;

    b[5] -= M(5, 6) * b[6]; b[5] -= M(5, 7) * b[7]; b[5] /= M(5, 5);
    interp += b[5] * (X[5] - X[N]).zw;

    b[4] -= M(4, 5) * b[5]; b[4] -= M(4, 6) * b[6]; b[4] -= M(4, 7) * b[7]; b[4] /= M(4, 4);
    interp += b[4] * (X[4] - X[N]).zw;

    b[3] -= M(3, 4) * b[4]; b[3] -= M(3, 5) * b[5]; b[3] -= M(3, 6) * b[6]; b[3] -= M(3, 7) * b[7]; b[3] /= M(3, 3);
    interp += b[3] * (X[3] - X[N]).zw;

    b[2] -= M(2, 3) * b[3]; b[2] -= M(2, 4) * b[4]; b[2] -= M(2, 5) * b[5]; b[2] -= M(2, 6) * b[6]; b[2] -= M(2, 7) * b[7]; b[2] /= M(2, 2);
    interp += b[2] * (X[2] - X[N]).zw;

    b[1] -= M(1, 2) * b[2]; b[1] -= M(1, 3) * b[3]; b[1] -= M(1, 4) * b[4]; b[1] -= M(1, 5) * b[5]; b[1] -= M(1, 6) * b[6]; b[1] -= M(1, 7) * b[7]; b[1] /= M(1, 1);
    interp += b[1] * (X[1] - X[N]).zw;

    b[0] -= M(0, 1) * b[1]; b[0] -= M(0, 2) * b[2]; b[0] -= M(0, 3) * b[3]; b[0] -= M(0, 4) * b[4]; b[0] -= M(0, 5) * b[5]; b[0] -= M(0, 6) * b[6]; b[0] -= M(0, 7) * b[7]; b[0] /= M(0, 0);
    interp += b[0] * (X[0] - X[N]).zw;

    return interp.xyxy;
}

With 4:2:2 the 1st hook simply copies LUMA to LOWRES_Y. You can remove it, but then you also need to rename all LOWRES_Y to LUMA in the 2nd hook (except line //!SAVE LOWRES_Y).

Hi igv, I have noticed that some of your FSRCNN models produce slight luminance changes that are barely visible when applied to luma, but can become quite obvious in other situations. This affects up to four of the models, the most affected is FSRCNN_x2_r2_8-0-2.glsl. Issues are disabled in the repository, if you contact me I could give you more details that I don't consider appropriate to discuss here.

FSRCNNX is latest and only supported version. Old models produce slight luminance changes because they were trained with SSIM loss.

hey igv, do you mind if you reupload the image comparison link at the top? it's not working anymore...

l am having a bad result compare to madvr's Bilateral sharp when applied to these stroke-like image which also common in anime.

Original:

madvr(left) mpv(right), I also saw this "red block inside of the stroke" issue in some anime scenes before:

madvr(left) mpv(right):

Doesn't look much different than spline36 from mpv (and without jpeg), just sharper.

Also these yellowish and blue strokes are part of the original image

Chroma bleeding sometimes happens with this shader, but FWIW it's mostly with a low bit-rate / quality images.

I didn't realize that those things are part of the original image LOL. Looks like KrigBilateral is trying to approach the original (also colors such as purple of the image above) and madvr's Bilateral sharp is trying to please human eyes IMO.

It seems krig cannot handle well when --video-rotate=90/270 is set in mpv.conf

Are you using hwdec? For me it happens only with hwdec and deband=no.

Yes, using"--hwdec=auto-copy". Except bilinear/oversample/bicubic_fast, there is no other cscale filters can perfectly show the rotated videos.

It's all mpv bugs.

@igv: Could you change bind order, as in the diff below?

@@ -14,8 +14,8 @@
 // License along with this library.

 //!HOOK CHROMA
-//!BIND HOOKED
 //!BIND LUMA
+//!BIND HOOKED
 //!SAVE LOWRES_Y
 //!WIDTH LUMA.w
 //!WHEN CHROMA.w LUMA.w <
@@ -52,8 +52,8 @@ vec4 hook() {
 }

 //!HOOK CHROMA
-//!BIND HOOKED
 //!BIND LOWRES_Y
+//!BIND HOOKED
 //!SAVE LOWRES_Y
 //!WHEN CHROMA.w LUMA.w <
 //!DESC KrigBilateral Downscaling Y pass 2

What happens is that in fragment shader CHROMA_pos outputted from vertex shared is not used. And during the translation/optimization GLSL->SPIR-V->HLSL the actual input of the fragment shader is removed, which triggers validation error:

ID3D11DeviceContext::Draw: Vertex Shader - Pixel Shader linkage error: Signatures between stages are incompatible. Semantic 'TEXCOORD' of the input stage has a hardware register component mask that is not a subset of the output of the previous stage.

Basically we have this:
vertex:

struct SPIRV_Cross_Output
{
    float2 _9 : TEXCOORD0;
    float2 _13 : TEXCOORD1;
    float4 gl_Position : SV_Position;
};

fragment:

struct SPIRV_Cross_Input
{
    float2 _17 : TEXCOORD1;
};

and by changing the order, we use TEXCOORD0 and it is not tripping the validation. We can discard things from the end of input list, but not from beginning/middle.

Hope I make sense, just small workaround to make it work better when validation is enabled.