Why would you need to scale images, rather than scale the entity that shows the final texture?
Software image scaling can offer much better quality then the bilinear hardware scaling fo sprite.
I run into the same problem. I wanted to have clear card images, with text and tiny details in them. Bilinear scaling made them look really bad.
I've used imagesampler
http://code.google.com/p/imageresampler/ with modified input-output routines, as it's opensource and has no dependiencies.
Other libraries tend to pull a bunch of dependencies, and I don't want to link directly to libpngm libjpeg, when the SFML is linked to them anyway.
Using it however takes some overhead to convert data between SFML internal format and that supported by imagesampler.
I can paste examples, but you can see them yourself when you scale eg. a newspaper page, first with HW bilinear, and then with ImageMagic using cubic or lanczos.
Real-life usecase: As I scale the card image size to the screen size I resize them by software scaling, before using them as 1:1 sprites.