I've been studying the avisynth branch of the Bifrost filter. Even though it doesn't use a sophisticated motion estimation algorithm (e.g., optical flow) I've found that it does a better job than pretty much every other software 3DCF* out there currently, TComb included.

I've been torture-testing it too. At reasonable settings it's hard to get it to artifact. It has an altclip parameter which gives you the ability to provide a different clip as an argument for the blocks it classifies as being "in motion". You can use this to process only the motion blocks with a different filter; e.g., the low-pass filter I've mentioned in other threads.

I've been using:

c.Bifrost(altclip=c.Spline16Resize(180, 480).Spline16Resize(720, 480))

This allows me to 3DCF static blocks and LPF motion blocks only. It works really well for the footage I'm working with at the moment.

You can considerably improve its motion detection by getting rid of all dot crawl first since it uses the luma plane to determine if a block is in motion or not and dot crawl screws that up.

I suggest capturing a S&W zone plate with your setup in order to find the best values to provide as arguments for the luma_thresh and variation parameters.

(* Keep in mind that this is for chroma only. You'll have to go elsewhere for dot crawl.)