Thank you a lot for the tip 
Hi, I think the solution could be easier.
The problem could be in the render integrator options. These are the options used in blender tests 2.61 (http://www.blender.org/development/release-logs/blender-261/blender-261-demo-files/):
And this is a simple scene using default glass (rgb(1,1,1)), a hdri image and an area lamp. It has 2000 passes and I am using blender 2.62:
There are just a few fireflakes. If you want to reduce them, try reducing glass rgb color to (0.8,0.8,0.8)
I have made tests with glossy shaders and with 2000 passes they looks perfect.
There’s a non-working attachment, was that the comparison? Anyways, looks well improved with those settings, cool. Can’t wait for future Cycles development, i’m already loving it even without big features such as well working displace ecc…
Is there really no “common” image processing filter that senses stray one-pixel noise and replaces is with a blend of the surrounding pixels? Compared to other filtering tricks, and certainly compared to optical flow analysis stuff, it seems like a relatively straightforward trick.
Also, can’t comparisons between different sample levels be used somehow? That’s a bit like the “different seeds” method, but without re-rendering the whole thing.
A bit off-topic, I’d love a Cycles setting to save out images at specified sample levels. I do it manually sometimes to analyze the sampling progression, and it can give some really useful insights about the renders. It would be great to have that available for those renders left cooking over night…
Great Tip! Thank you a lot 