Guiding laplacians to restore clipped highlights

hanatos · December 30, 2021, 1:47pm

it’s pretty standard pull/push (maybe closest to this if you want a paper)

and for the colour reconstruction step i’m using the minimum channel of the three rgb channels to guide the others by ratio determined by pull/push. only that this guide-by-minimum did not really work perfect in all cases… when the minimum changes over the course of the blown highlight. so i came up with a horrific formula that represents a smooth blend between the minimum method:

github.com

hanatos/vkdt/blob/master/src/pipe/modules/hilite/assemble.comp#L120

    
      
          #if 1
            // rescale upsampled values to match unclipped values in fine buffer
            // this is risky because min might be close to black/0 so it may be unstable:
            // guess channel by darkest blurry version
            float sr = fine.r/max(0.001, upsm.r);
            float sg = fine.g/max(0.001, upsm.g);
            float sb = fine.b/max(0.001, upsm.b);
            float wr = exp(upsm.r - max(upsm.g, upsm.b));
            float wg = exp(upsm.g - max(upsm.r, upsm.b));
            float wb = exp(upsm.b - max(upsm.r, upsm.g));
            float scale = (sr * wr + sg * wg + sb * wb) / (wr + wg + wb);
          
          
  vec3 rec = upsm * scale;
          
          
  // try and only apply if we are reasonably sure it's not an edge but a gradient:
            // (i think this goes the right direction but maybe t needs to be more
            // aggressive, i.e. degenerate to the above smooth with t=1 sooner)
            float t = 0.0f;
            if(any(greaterThanEqual(fine, vec3(white)))) t = 1.0f;
            // else t = clamp(1.0-0.3*length(fine-upsm), 0, 1);
            else t = clamp(params.soft-0.3*length(fine-upsm), 0, 1);

and (as you can see a few lines below that) there’s another parameter that determines whether you’d like to prefer the low frequency or the high frequency scales if you have a choice (i.e. both are good averages without blown values).

it’s more on the blackmagic side of things, but then again this is an ill-posed problem, and at least it runs fast.