I tried using MMT to extract the very highest scale length permitted, at 256, while disabling all lower levels. I applied that to an RGB image, and the result still shows a lot of small-scale red structure. When I split out the RGB channels and do them individually, there is no problem. All that remains is large scale structure in each of the planes.
It appears as though channel 0 processing halts around scale lengths of 2-4. The remaining channels are handled correctly.
Example shows a region measuring about 1000 pixels square.
