I’ve tried my theory of bending the normal. Currently, I’m just creating a dot product from the camera to surface point direction and the vertex normal. And use that with an eyeballed curve to bend the normal towards the camera.
Example with a light shining towards the camera and roughness values of 1 and 0.25.
It’s less pronounced with a light behind the camera but still a clear improvement (only roughness 1 here)
