Multiscale algorithms (wavelets et al) are, by their very nature, scale-dependent. This means that sometimes a preview may not work as expected, simply because the whole image contains structures at scales that are not well represented in the preview. The deringing and wavelet regularization algorithms implemented in the Deconvolution tool use multiscale techniques, so this problem can happen.
When this happens, try to use a larger preview covering a wider range of image structures, for example more background and nebular regions. This usually provides accurate results. For maximum accuracy, define a preview covering a relatively large area representative of the entire image.
In some multiscale tools, as for example MultiscaleLinearTransform and MultiscaleMedianTransform, we have implemented quite complex algorithms that precompute the transforms for the whole image and then use a cache to accelerate subsequent operations. This provides completely accurate previews (including real-time preview). Unfortunately, this is not possible for other tools such as Deconvolution and HDRMultiscaleTransform.
