How much experience do you have with UE4? This is a bit complex.
The basics of this would require 4 render targets. 2 to store position and 2 to store velocity. Each tick, one if each is swapped for the other as a texture parameter, in 2 dynamic material intances, one which calculates position, the other velocity. These dynamic material instances would then be drawn to the render targets not being sampled.
You would input vector parameters for the velocity and position of a sphere you control. Calculate the velocity by multiplying the velocity of the sphere by a scaled distance between the sphere position and your position render target sample. Add another velocity component of the difference between the original point cloud texture and the position sample, multiplied by some scaled distance between them. Add that to a sample of the current velocity render target to get your final velocity to draw to render target.
In your position dynamic material instance, add a velocity sample times a scalar parameter for tick delta time to a position sample.
The bensacon cloud is almost certainly too big for this, unless your card is a monster. I’d start with a 512x512 cloud texture and render targets. Also follow some of RyanB’s optimizations in that thread. If you pair precalculated uvs with your vertices in a static mesh derived from a procedural mesh, you can save quite a few instructions in your material.
Also, instead of quads, if you use equilateral triangles centered on (x,0,0) and flip their vertices over z with world position offset, at about 60 hz, they look pretty good with temporal AA, and you halve the number of polys you need to render.
Ps, I only found this randomly, next time you should probably ask your question in the thread, where people know how to use my technique.