Hi David,
This is really weird. ATWT is a complex process that cannot provide very accurate downsampled real-time previews (just orientative ones), but this kind of divergence is new to me. The real-time preview works accurately for intensity transformations and relatively simple algorithms such as unsharp mask, convolutions and morphological filters with simple structuring elements. For more complex tools and processes the RTP is less accurate when the previewed image has been reduced (subsampled).
I'll investigate this. As a general rule, for complex multiscale processes we recommend using previews instead of the RTP. There are also some multiscale processes —notably HDRWT— that don't provide consistent results for partial previews (a partial preview covers less than the whole image).
