Hi Chris,
Patrick has set up a nice series:
http://www.cineversity.com/vidplaytut/global_illumination_part_01
This question comes from time to time. I would avoid presets, they keep you in an average state, always. Each scene is different.
What I found, and each thread brings something new to that theme, that the set up of a scene is certainly highly influential.
To have everything generating and receiving GI in scene is certainly not a good idea, e.g., light source have mostly no benefit from receiving nor has a dark object a big influence while creating. (Illumination channel/tab>Material editor). Typically a plant object with a huge number of polygons will only slow down the calculation, if included, but has little influence to the result, sometimes it makes sense to replace objects (not visible to the camera but visible to the GI, and vice versa)
To specify portals and gi light sources in the same tab is absolutely helping.
Small sun lights in global illumination HDRIs are better painted out and replaced with a disc and a bright material.
What ever object can be baked will limit the flickering, sometimes camera projection will do as well.
I found that backing a Chinese light close to the camera can help. It is a sphere that bakes all receiving light. Sometimes the camera move is pretty simple forward, which allows to just bake a simple shape of the scene with the light behind the camera, a much faster rendering and a more stable scene is the result. Replacement as with the plant example is key for me at least.
Tangential received light is mostly less stable than perpendicular.
I place sometimes a cube or sphere around the scene with a black or dark gray non receiving/generating material around the scene. Just so the area is limited, which is of course case specific.
Light objects instead of Luminance material takes one bounce out of the equation and provides a different shadow.
Objects specified as light sources with luminance material should have their surface fully illuminated, partially darkened with a texture will mislead the “rays”, and the results are more than bad.
Intersecting models might cause problems, clean modeling is never a bad idea.
GI is not a solution for a patchwork scene. I encourage to know what you like to have and get closer with supporting lights in the first place.
Last week We discussed this here. http://www.cineversity.com/forums/viewthread/1490/
To my surprise the secondary method that solved it was QMC. Never thought about that, as QMC is typically the noise producer.
In some cases I had good results by storing the IR cache while exploring the scene with a different camera, to gain more data. Where ever I had a problem render to the picture viewer a detail view and store more information there.
There is certainly more, but I’m sure, going through Patrick’s series will “illuminate” this option much more.
The main key to success is certainly to train your eye to “read” the dots in the pre-pass, as it gives a relative good idea where to little information is produced.
After all, and I know I repeat myself here, a clean set up of the scene is mandatory for good results (I haven’t seen yours so this is just one point on my typical list)
All the best
Sassi