I ended with GI/AA - 64 and LSM - 0,5. Is that a good idea?
It might be. Could you post some render and viewport?
Buckets with adaptivity pay off when you have large flat areas that refine fast. So when you are rendering without adaptivity, Corona will sample those good pixels anyway wasting time. In hard noisy situations adaptivity may even slow down rendering because it will add lots of samples in those places. At least that's how I think it works. :)
Unfortunately, there isn't single flat area in my render. I did some tests yesterday with 3/3, 2/4 and 1/5 settings for adaptivity and contrary to my expectations, settings with more initial samples and less adaptive passes rendered fastest. So you might be right, maybe my case does not suit for adaptive sampling very well.
Anyway, if you use a lot of direct lighting then I would try using default 16 for GI and some higher values for LSM (like 6 or 10 maybe) not sure about sss sampling though. I guess default 16 for gi should be ok.
It would help if you could post a screenshot of your scene setup and maybe also some screenshot from rendering.
If im not sure where is most noise coming from I disable both gi solvers to see how much direct lighting really contributes to the scene and if most noise is from GI or direct light.
Direct light is cleaning very fast, most of the noise is coming from translucency. That's why i reduced LSM.
Anyway, i rendered test sequence over night with bucket engine. Settings were GI/AA - 16, LSM - 0,5, Initial samples - 3, Passes - 3. With some denoiser treatment in AE it came out quite acceptible.
p.s. sorry for lack of screenshots, but i don't think they woul be useful in any way. It's just single object in open space - there's almost no indirect. Only direct, reflection and translucency.