Skip to content

Instantly share code, notes, and snippets.

@igv
Last active March 22, 2024 03:04
Show Gist options
  • Star 120 You must be signed in to star a gist
  • Fork 4 You must be signed in to fork a gist
  • Save igv/a015fc885d5c22e6891820ad89555637 to your computer and use it in GitHub Desktop.
Save igv/a015fc885d5c22e6891820ad89555637 to your computer and use it in GitHub Desktop.
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 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;
}
@Alexkral
Copy link

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.

@igv
Copy link
Author

igv commented Jan 16, 2020

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

@arennf
Copy link

arennf commented May 2, 2020

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

@VEWION
Copy link

VEWION commented Dec 22, 2020

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:
original

madvr(left) mpv(right), I also saw this "red block inside of the stroke" issue in some anime scenes before:
madvr_crop1 mpv_crop1

madvr(left) mpv(right):
madvr_crop2 mpv_crop2

@igv
Copy link
Author

igv commented Dec 22, 2020

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
orig-zoom

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

@VEWION
Copy link

VEWION commented Dec 23, 2020

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.

@hooke007
Copy link

hooke007 commented Feb 17, 2021

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

@igv
Copy link
Author

igv commented Feb 17, 2021

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

@hooke007
Copy link

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

@igv
Copy link
Author

igv commented Feb 18, 2021

It's all mpv bugs.

@kasper93
Copy link

@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.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment