Indeed, since I posted that comment I've been looking into it further. I read on wikipedia [0] that: The DCT based form cannot guarantee that encoder input wou

Indeed, since I posted that comment I've been looking into it further. I read on wikipedia [0] that: The DCT based form cannot guarantee that encoder input would exactly match decoder output since the Inverse DCT is not rigorously defined I don't get that - I thought the inverse of the DCT was just the linear combination of the components. So...I've been playing around with it myself to see what happens (using graphicsmagick). I converted an image to jpg (quality 70) and then bmp -> jpg again. There are definitely some minor artefacts. I then iterated that step jpg -> bmp -> jpg (quality 70) a dozen times. These further iterations don't add any additional artefacts. So it sort of gets into a stable state. EDIT: reading that quote from wikipedia again it looks as though it's just saying that jpeg is lossless (ie, we apply quantisation). [0] http://en.wikipedia.org/wiki/Lossless_JPEG http://en.wikipedia.org/wiki/Lossless_JPEG