Layers and layer groups are re-rendered in Photoshop at the requested resolution before we create an asset for the requested resolution (see question below for how we render smart objects).
Assets requested from Vector layers are scaled in vector space before they are re-rendered at the requested resolution. If a raster layer or layer group is smaller than the requested resolution we upsample it in Photoshop using the Image Interpolation algorithm set in Preferences > General > Image Interpolation (the default is Bicubic Automatic).
Layers or Layer groups that are being output as SVG are resolution-independent if they contain vector shapes. If the SVG contains rasterized images (which are included inline using base-64 encoding) they are not resolution-independent. Currently these images are not scaled up for higher resolutions, but this is a bug we hope to fix with the next dot release.
Have you seen this thread?
It seems to describe a reproducable difference in the handling of placed and converted pixel Smart Objects in Generate > Image Assets.
Is there an (easy) explanation for this?