Temporal stability if you read the white paper is what this technique does well. It’s a single pass system (basic 1 bounce) that uses past frame and predictive evaluation based on velocity of geometry per pixel to be able to use data to represent multi bounce indirect with no extra cost. Yeah it’s not going to be perfect because you don’t have all scene data for eval (unlike offline renderers that have to have all geo) but has enough through temporal remapping and view point guard banding (e.g rendering view at larger size by 10-20% for example) based on the fact human perception can’t correctly interpret indirect lighting and reflection beyond our viewing point per frame. By evaluation of this extra perception data (even if not from a 360 degree point of view is 90% of the time enough for realistic perception of lighting and reflection) you get a near 99% acceptance of the result.
Also you should look into what i posted for devs to look at not long ago, SVOGI even with intels steps are still not great, ive been banging on about SV DAGS for months. Check this Budd Compact Precomputed Voxelized Shadows . Great papers http://www.cse.chalmers.se/~d00sint/