Hummel, today I really considered to use normals derived by the same exact method you mentioned instead of sampling the gouraud shaded ones, but as you already said, this leads to hard normals. But if I tweak the blur, it could work for the upscaling.

Question: if I bilinearily downsample the depth image and create worldspace normals from the downsampled perpixel depth values, do I get the same normals if I would have create the perpixel normals from the hires depth values which I would then have downsampled, too???

That would be great, because I could save lots of memory then (and processing time, pushing it down the very long throat of your videocard... gag!).