As the article says, transformation is best if your game mostly uses primitives (lines, rectangles...)
From looking a screenshot you may think it's also better visual quality for sprite, but in an actual game it will cause graphical artifacts which are sometimes jarring.
The issue accurs when several sprite bitmaps are supposed to be exactly side-to-side (sokoban player pushing a crate) or overlayed ("damaged variant" of a building, landscape composed of several layered images...) There will be a 1-pixel inaccuracy, and it's very visible if things are scrolling/rotating/zooming as the elements will "float" within a 1-pixel distance of the ideal position, instead of staying stuck together.

Stretched buffer doesn't have this issue because integers match exactly the coordinates and sizes of elements in the original referential, so the computations are pixel-perfect.

3D rendering systems don't have this issue because the system will keep track of all coordinates with decimal precision, and takes the entire scene into account when rasterizing.

Edgar, I tried adding these flags, however, the image looks more blurry then but not as nice as with the transformation (see attachement #2). Maybe there is a hidden filter applied when transformations are used?

Audric, I did not know that. In fact I found for example this thread where it sounds like people actually prefer transformations over a stretched buffer because it is less GPU/memory intense because the GPU will do scaling anyway. But if you say that this could yield in a 1-pixel inaccuracy this would be a big disadvantage (at least for me)

OK guys thanks for your help. I found that pixel inaccuracy may occur indeed (see attachment #3) when using transformation (especially non-integer scaling) so I think I'm gonna stick with the stretched buffer...

OK this makes sense indeed...

I'll try to figure out where that 'inaccuracy' may come from...