Direct port from Sirikata's C++ Jpeg<->ARHC library into a standalone tool. When you compress a jpeg file with a standard tool like bzip or xz or lzma, the file size often increases, but certainly does not decrease much. Sometimes you have a lot of images you want to archive, but you don't wish to alter the input source at all. For those cases, this tool can bring the size of the source file down by 10%. In some cases you want to make sure the pixels stay exactly the same color but that the source file itself may be altered. For that, a tool like jpegrescan may be more useful https://github.com/kud/jpegrescan The way this tool operates is that it takes a pass over a jpeg file. When it encounters run-length encoded items, it places them in a single, difficult-to-compress "bitbuffer" area. As it finds Huffman Coded components, it simply huffman decodes the items. Each item is placed in an array corresponding to its DCT component number and color channel index. The arrays are stored in zigzag order towards the beginning of the file, one channel at a time. To reconstitute the original jpeg, the file is read, and the huffman-coded items are pulled from the appropriate section in the beginning and re-encoded with the original huffman tables. The bitvector copied to the corresponding areas, interspersed with the huffman-coded items. When the huffman data values are in the decoded form and in a series of arrays at the beginning of the file, the jpeg is able to be efficiently compressed by standard compression tools. This brings the final file size down ~10% below the original and allows the original to be reencoded.