You can easily quantify the variance and do standard model-comparison/hypothesis-testing if you want statistical-significance levels. For many datasets these days, this is hardly even a consideration: even a 1% compression improvement is clear.