[mtripoli]

When rendering I typically let it cook until I'm happy with it (or a dead line is looming). So my question is, how long is long enough? I've occasionally played with rendering gem stones for fun and it seems like they will never clear of all the little spectral noise dots. What if they ran for a month, a year, 10 years? Does a render (forget the gem sample) continuously get better, or is there a point where it doesn't or can't get any better? How does one know? Is there a quantitative way to know, and if so could there be a little box that says "This render will be BEST finished in XX time" or given the psychology of people would that freak them out to see it may take 2 years to reach "done"?

Just thinking out loud...

[polynurb]

some graphical/statistical indicators of how complex the scenes calculations are and how far their progress in solving is would be extremely helpful..

so thinking out loud as well.. eg. FIRE changes priority upon the different phases of calculation, so there is "awareness" inside the maxwell engine about what the components of the image calculation are.
as far as we can guess these are at least:

direct lighting
indirect lighting
direct reflection caustics
direct refraction caustics
indirect reflection caustics
indirect refraction caustics

the engine however will never know what is important in to us in the image we render, but it could still give us statistical data to what extend X% of pixels are dependent on each of these calculations.

eg. in an daylight interior scene if we do not leave some windows open without glass geometry, there would be an "alarm bell" showing up, indicating that the gross amount of light power composing the image is of caustic nature.
somesort of a scene debugger...

i think it would be very interesting if we could interact more closely with the core of maxwell.
assuming that if the user could give feedback to the engine telling it what parts of the image are of greater importance, the image could look better to the eye at lower SL?!
a tool showing the areas of eg. multiple reflected indirect caustics prior to rendering could then be used to manually mask these areas and so telling the engine to emphasize them.

[Bubbaloo]

the user could give feedback to the engine telling it what parts of the image are of greater importance

Wouldn't that make Maxwell a biased engine?

[polynurb]

not really... the core itself would not need to work differently, in terms of the math it uses.
it would just spent time differently.. when the highly diffuse surfaces in the image are clean@sl 12, power is directed at other parts.

so i'd say more tweakable. right now the decisions how an image is converging are taken outside of the user's influence.
which, i agree, is a part of what gives maxwell it's reputation in image quality.
on the other hand, seeing maxwell go more and more into film productions, a more customizable approach could be of interest in order to overcome certain problematic situations.

but hey.. i have no idea how a 2d representation/masking etc. could be fed back into a 3d space calculation, pathtracing thing.. just sketching some thoughts here