Frenetic Pony
Veteran
Not sure how much restir can halp with such a low sample count. Also, i guess they already have prefiltered / low res geoemtry, so an approximation of cone tracing, which already gives them a better inital condition, and making point sampling more effective isn't their sapproach here. But really not sure.
If you ask me, their reflection results already are acceptable.
It's just that everybody is a bit spoiled from HW raytracing results, which have their price, but can be used as well where applicable.
Personally i'll enjoy the improvement over SSR and cube map hacks. (If frame rate is high enough for joy.)
For a game like Spiderman, a single planar reflection plane for the building in focus, plus some blending to alternatives for the rest wouldn't be that bad either, see Hitman.
Should be able to help with a good blue noise vector sampling scheme/taa. You treat your previous reflection shadow ray as valid next frame if nothing moves and re-apply the shading, while you get a new shadow ray to test your new jittered reflection vector this frame. As long as both the primary surface and reflected sample are static you build up samples over time, higher res reflections even on glossy/mirror surfaces. And of course spatial sharing helps even more with rougher reflections/diffuse.
One scheme I've been interested in is vastly increase screenspace shadow rays. You begin sampling rays in SS anyway, quadruple that to 1/4th sample rate and then only continue 1/4 of those samples out to HW tracing once they miss.