JPEG compression 600 times over

Generation Loss from hadto on Vimeo.

The pitch for this little video has the elegant and demented simplicity that is the hallmark of all great ideas: "Open the last saved jpeg image. Save it as a new jpeg image with slightly more compression. Repeat 600 times"

Generation Loss (via Beyond the Beyond)


  1. gOSH … that’s not so bad at all. My Sony Mavica camera (with a mighty 0.300 Megapixels) used to do that much damage to a scene in *one* jpeg compression!

    Happily, tech has advanced so that RAW/TIFF formats can be supported … no compression!

  2. If this were a virtual Rorschach test, that stuck me as the plaques, tangles, and misshapes of the affected Alzheimer’s brain.

  3. 36 Jphilby: I know exactly what you mean, I still have one of those old Sony Mavica around here somewhere. And to think they meant you to use a 3 1/2 floppy in the camera. Even at the time, I thought it was really awkward.

  4. @#4 Takuan – that’d take 666 compressions. With any luck you’d come up with a Jesus or Mary eBay bonanza image before that. Start with a grilled cheese pic :-)

  5. This should also be a lesson to people in the difference between lossy and lossless compression. (Or why you shouldn’t use JPG for crisp, spot-colour images like logos.)

  6. #14 – it would be great for that if they didn’t increase the compression each time they resaved it. it’ll still get the idea across, but they cheated.

  7. #3 KAINI

    I really liked that! Compression, you say? It seemed that looking at the results there are definitely some very trackable markers in that image. I’m interested in seeing from what they drop out of.

  8. #9 Ãœbertragung, now I’m going to have to try that, you big jerk.

    Of course, it will take a clever input source, because an mp3 has the time dimension.

    Om? Famous statement? Famous sung phrase? Opening chord to Help? Tell me what you want what you really really want? It’s got to be short, iconic, and demonstrate the artifacting in an interesting way.

    It’s a tougher choice and if it’s too long the effect won’t work, and hurrying up the effect by crappy compression could ruin it as well.

  9. Can someone remind me why the final image has a smaller file size? I took the advanced mathematical theory in college- but, it’s been (more than a) few year, and I’m kind of a glutton for math theory.

  10. #16: i don’t know what sort of video compression was used. it looks like the technique employed was to deliberately drop keyframes all over the bloody place. one thing the maker did comment on is the fact that different codecs produce different results. avi is ‘glitchy’ whilst wmv produces ‘more organic results’ (probably the only thing wmv is good for).

  11. @Anon 18,

    Something short, iconic, and demonstrates the compression in an interesting way?

    How about Sean Connery saying “One ping only” followed by the ping from “The Hunt for Red October?”

  12. Nifty. I once did something along these lines for an animated film–I “zoomed in” on the compression fuzz by cropping an image, resizing it to the original size, saving as jpeg, and repeating ad nauseum. The artifacts assumed their own substance as apparent objects in the “zoom.”

  13. Re: MP3 equivalent- Instead, just take a medium song (say, 5ish minutes) with a reasonable amount of repetition (say, anything by U2), perform the repeat compressions. (MP3->WAV->MP3, &c) and gradually fade to the more highly compressed version as the song progresses.

  14. I swear that’s one of the evilest things Ive ever seen. They should put that in the dictionary next to Entropy.

  15. I do hope they did this with a shell script, or something similar, and not by manually opening and saving a jpeg in some GUI software 600 times over.

    @Daemon: Didn’t notice that. It’s a bit silly – the effect would be the same if they left the compression constant, though it would take longer. (That would be a more controlled experiment, too.)

  16. This is NOT generation loss. This is increased loss due to increased compression. I would expect the same result when saving the original image 600 times with increasing compression level.

  17. @ghstpg: perfect.

    @Anon, et al: Here’s an MP3 attempt FWIW-

    The windows XP start theme iteratively degraded 42 times (666 times was way too long, 42 the next best number):

    It has struck me that this could be an evil prank to play on someone… switching with the usual wav file. A startup that never ends.

    (An odd artifact from using the LAME backend for encoding was a gradual increase of silent padding fore and aft of the original waveform…no clue why this occurred, but it winds up lending a sense of hope that it is over, which then is crushed by another, more degraded rendition of the sound. enjoy.)

  18. Prior art:

    The original ‘Doctor Who’ titles had a weird-looking set of blobs that moved out from the center of the screen. That was ‘video howl-round’ – you pointed a TV camera at a TV screen, and vary carefully twiddled the gain. The usual result is either black or white, but it is possible to get a set of randomly moving blobs.

    I think Douglas (oh it’s not Martin Gardner) Hofstader refers to something similar in Metamagical Themas.

    I always meant to try the same sort of thing using a photocopier.

    You have to change the JPEG compression, or shift the origin, or the image does not continue degrading.

  19. Well, if you repeated the compression of an unaltered image with the same program at the same compression, you may not see any degradation at all.

    Depends on the code, but I’d say it’s more likely than not you won’t see any.

  20. This would have been a great way to explain lossy compression back at school. I am forwarding this to my old professors.

  21. I did this when the JPEG standard first came out, but as a series of three images printed in a magazine article. While JPEG artifacts are annoying at just about any level, for casual viewing they really aren’t much of a problem until you get to very high rates of compression, or if you have details that need to be very clear, such as small text.

  22. Uh, no, wait a minute. When you increase the compression, you decrease the number of available states, so you decrease the entropy. The over-compressed images should look less detailed and less random. Something is amiss in my understanding.

  23. #42: you’re looking at it wrong. Yes you decrease the information content, but jpeg uses a DFT so the basis of information is a (co)sine series. So look closely and see how the picture turns into basically a superposition of a few 2-D sinusoids.

  24. I did something like this with photocopiers as an art student, back in the mid 80’s.

    The analog nature of the process produced MUCH more interesting results!

    Re-Photocopy something enough times and you end up with something that looks like sand dunes.

Comments are closed.