Quaternions are in a sense just a 4D vector, so you can perfectly calculate the average component wise. Once you renormalize, it should be a rotation again. Of course, if the different quaternions are very different (say almost opposite of each other), the numerical errors are big, but the same is true for every way to calculate an average.
But keep in minde that my knowledge of quaternions for 3D rotations is 30 years old and purely theoretical, so it's just a thought, not an advise from a pro in that area.
Thank you! So add up the 4 components separately, divide by 6, and renormalize back into a vector4? I'll try that and see what happens!
That could help me get rid of that costly 6x Slerp loop.