I am curious on what technique you are using. I am trying around with some depthbased and also with some normal based stuff, but I am not happy with both of them... Not even with their combination wink.
I also toyed around with a combination of scene-normals and scene-depth but always got strange results. When using the actual model normalmaps it works as it should:
As for the technique: The depthmap gets refracted by a noisemap and then the original depthmap gets substracted from the blurred one. I do this like 32 times (each time with slightly different texcoords and tilingfactors) and then divide the result by 32 . I think that's basically it. If you are interisted, i can send you a copy of the shader sometime next week. Maybe we can work something out together on how to correctly use the scene-normals to create a more realistic look
