>>  1. If you really want to compute three Gaussian noise values per pixel,
>>     you should probably use the Ziggurat method for generating the random
>>     values; the ratio method is very slow. (That, and you should not
>>     do a divide per nextDouble()!)
>>  2. There's absolutely no point in counting downwards in a loop that doesn't
>>     use its index; it doesn't buy any speed, and has not since sometime
>>     during the nineties. The compiler is able to do such transformations if
>>     it is a win.
> "rgbnoise" is a straight port from gimp. We drop it for now if
> consensus is that it is too slow.