unfortunately, prediction assumes one decodes everything, since the spatial
predictors are located on the boundary of the 16x16 or 4x4 blocks. This is the
downside of using spatial prediction.
Still, doing on-the-fly resize reduces the CPU usage because there's less
yuv->rgb conversion operations to perform. And you can bypass the in-loop
filtering process if the reduction factor is important (for instance when you're
only extracting a thumbnail out of a large picture).
But most of all, the reduction in memory usage if the interesting part of
on-the-fly macroblock transformation.