Quote - As pjz99 said, the values depend a lot on what's in your scene, but I can at least tell you which way the sliders go for "better".

First (raytrace) Bounces: Same as Bounces used to be - limits the number of Reflective/Refractive/Transparent surfaces hit. Needs to be at least 1 for IDL. Higher numbers give more accurate optics, but at increased render time. I haven't needed more than 1 so far.

First (AO) Irradiance Cache: Used for AO and disabled for IDL. If you're doing IDL, leave this at 0.

Irradiance Sample Size: Globally reduces the maximum distance allowed between IDL evaluation points. (I'd guess it does something similar for AO when IDL is off.) Lower numbers mean more evaluation points, smoother IDL effects and substanially more render time. I've always kept it as the top end of 20 so far. My reccomendation would be to lower this from 20 only as a last resort**.**

Intensity: Set the global balance between IDL and direct/specular lighting. I've always used 1 so far.

Now we hit the most important ones for IDL.

Bounces: Sets the number of bounces off surfaces used to sample where IDL is coming from. Depends on how complex the scene is. Higher numbers allow light to sneak under tables and things. Luckily, because of the magic of the Irradiance Caching algorithm, the render time does not depend exponentially on this number. Starting with a value of 5 seems plausible to me and don't be too afraid of even higher numbers.

Samples: This sets the number of random rays sent out after the last bounce to figure out how much light is arriving at the last bounce point. It has to be pretty high for two reasons - one, it is a statistical sample, so the error in the result goes down rather slowly with the number of samples, and two, the rays have to sample a complete hemisphere at the last bounce point. If there are too few rays sent out, they might miss important nearby geometry. Expect this number to be several hundred, if not thousand.

Second (IDL) Irradiance Cache: This is the most critical number. It effectively sets the accuracy in IDL that you're willing to accept and will cause the render to churn until it gets there. On the other hand it is the source of much magic as it forces more IDL evaluation points precisely in those places where the geometry is changing quickly and fewer where not much is happening. The actual maximum error in percent = 100 - Irradiance Cache. This means several things. First, high accuracy is to high numbers, or to the right. The usable numbers must be less that 100; 100 leads to no caching and unbearable render times. For high accuracy, the slider range is very compressed at the top end. E.g., to go from 5% maximum error to 2.5%, the number goes from 99.5 to 99.75. This is hard to manipulate with the slider and you probably need to type it in. Also, for each cut in the error by half, the render time could as much as quadruple.

These last two settings are the heart of IDL and they are somewhat coupled. For example, setting (IDL) Irradiance Cache high for high accuracy will fail if Samples is not also high enough to get reliable IDL estimates at the IDL evaluation points and vice versa. Fortunately, the test renders I was doing seem to be very demanding since they have such flat, untextured surfaces. For most renders, quite a bit sloppier calculation of IDL is fine. For starters, try (IDL) Irradiance Cache at 50 and work your way up carefully.

I think that is the best I can say at this point. I hope it helps.

Whichway

Thanks for putting this all in one comprehensive post, I'll be printing it and referring to it often till I can feel comfortable with it. Render times on my old PC are insane with just the test object, and was getting no where quick with testing.