[t0rakka]

Guilty as charged.. the 1% was a napkin estimate.

The True Cost in this specific case is (0..7 bits to fill a byte) + 2 bytes for the RST marker for each interval. 320x240 image and RST interval for each 240/8 block means 30 RST's so something close to 75 bytes "wasted", or, other way to look at it is ~3 bits per row.

The reason is that for the JPEG the Huffman tree is fixed and is transmitted with the image data so restarting is cheap; it does not start from scratch and build a new tree each time; each decoded symbol is also independent so you can stop and restart at any time you want. I don't have any ANS experience at all so won't say anything about that.. but it would be nice if some sort of way would be possible to isolate the ranges.

Good observations, btw. :)