I think these end up looking about the same usually. The second method transforms and resamples the tensor data, and it's not really clear what the best way to do that is for tensor data, as opposed to the acquired DWI data. On the other hand, transforming the DWI data also squeezes and stretches the data, so that voxels are not exactly the same size as the original, and multiple voxels combine to make each voxel in the output. The interpolation scheme could have an important effect. One should be able to do all analysis in the original space of the subject and lose a little detail in transforming to a standard space because of the transformation and interpolation. The effect will be dependent on the DWI data, the fibers of interest, the standard space template and the transformations.