IMO STF is being given a bad scale estimate, which causes it to produce results not intended.
This doesn't happen. Each scale estimates is multiplied by the required constant to make it coherent with the standard distribution of a normal distribution. For example, MAD is multiplied by 1.4826, average deviation by 1.2533, and so on (anyway, in the version of Statistics that I'm going to upload later today, you can select between normalized and unnormalized scale estimates).
The "problem" here is a behavior change. In general, STF works now much better IMO. However, there are cases where the old automatic stretch functions can be desirable. For performance reasons, STF cannot use a too expensive scale estimator, such as Qn or IKSS. It now uses MAD to characterize the main histogram peak, while previous versions used AvgDev. This is the eternal battle between robustness and efficiency. I have favored robustness in this version because I think it is a more useful property for most raw linear images, but nothing can work ideally always.
Tomas, could you upload this comet image, so I can make some tests?
