Importance-based/guided adaptive sampling

Current adaptive sampling works by sampling a given pixel until a value (a function of variance and brightness) drops below a specified threshold. I propose an adaptive sampling mode which uses a an importance map, derived from the variance and brightness data of the rendered image, to allocate pixel samples probabilistically. Instead of having a maximum samples/pixel value, we can use an average samples/pixel value. Blender already importance samples sky textures to cast shadow rays, so it is entirely possible to importance sample an image.

This method has the following advantages:

• Minimum samples can be very low or even zero
• Probabilistic nature makes it easy for new high-variance pixels to be discovered (reduced under-sampling)
• The last samples of the render won't be focused exclusively on clearing up fireflies/high-variance spots (reduced over-sampling)
• Time limit rendering will work better with adaptive sampling
• Settings could be more intuitive, or even the same as without adaptive sampling
• Potentially greater image quality, as importance sampling is very effective at allocating samples in other domains of rendering (tests will be needed)

I've had this idea for over a year now, and I haven't been able to think of any major drawbacks or limitations that threshold-based adaptive sampling doesn't already have.

For example: how would threads be allocated when most samples are concentrated in a single region of the image? The same way as they are when most pixels have converged in the existing method.

I'm sure there are some specific scenes and edge-cases where the old method will work better though.