hi thetule
first, nice animation, storry.
Well when you render an inremental i-map, you should use incremental ad to current frame, which forces vray, only to render those areas and details, that come new to the viewfield of your camera.
at last you get a god overal lighting for the animation you did. well , it sounds as if you did do this at each single frame, which costs much rendertime and also isnt nessesair for a flightthroug.
use incremental ad to single frame, and go to your rendersettings, to force vray to render only about each tenth frame. this is enough if you make such smooth and soft animations like you did. Here you moreover can use higher settings as you would in single immages, because you only render each tenth frame and also only one time the whole immage, because the follwing immages only compute the new areas that com new to your camara.
Well now your problem.
for your problem, you cant use incremental ad to current frame, because like in real world, moving objects change the whole light situation in each frame, you have. because of this reason, each frame for itself has to get a whole new light simulation. only way to get this isnt incremental ad to current frame, but " Single Frame".
The problem with this method is the stochastical analysis of the light situation. Because modern Rendersystems reach their speed only of analythical proximations of the real lightdistribution in the scene, immages would slightly change in light from frame to frame. this is usual, and depends on the quality settings you take for your rendering.
The lower your settings are, the more the frames flicker from frame to frame.
you only can solve this with very very high settings. And a way to do this is using for example lightmap for secoundary bounces, and direct computation for first bounce. Well of course is the most timeconsuming but at last most perfect way to get perfect results.
Another way was lightmapping for secoundary bounces (only with vray 1.4570), and l-map with very very high settings.
At last all of this take much rendertime.
the third aproach was to compute l-map only for objects that doesn´t move. All moving objects get excluded from i-map calculation, and get renderted with fake radiosity.
An example is the animaiton of a medicin machine i did some tima ago. (you also can find the fakelight script there)
here´s the link. 
http://www.chaoticdimension.com/forums/phpBB2/viewtopic.php?t=7534&highlight=