How to losslessly encode a jpg image sequence to a video in ffmpeg?

Question

I have a large set of jpgs that I want to convert to a video losslessly (or, at the very least, very close to losslessly as long as the encoding time isn't much higher than otherwise).

Naively, I would think that there should be some codec that can store each individual jpg frames as-is (without recompression), and perhaps achieve some nice compression by replacing some of the frames by just the information on the delta from the previous frame. In my case there are many sequences of frames which are identical to each other, or that have a tiny difference between them.

Is there some codec and suitable settings for ffmpeg that can achieve this?

Answer 1

Just mux the images

You can simply mux the JPG images to make a video:

ffmpeg -framerate 30 -i input%03d.jpg -codec copy output.mkv

Note that if you omit -framerate then a default of -framerate 25 will be applied to the input.

Lossless optimization

You can use jpegtran to perform lossless optimization on each frame which may provide significant file size savings:

mkdir outputdir
for f in *.jpg; do jpegtran -optimize -copy none -perfect -v "$f" > "outputdir/$f"; done

Now mux with ffmpeg as shown above.

Checking that it is actually lossless

The framehash muxer can be used to compare the unique hash of each frame to ensure that the result is truly lossless:

$ ffmpeg -i input%03d.jpg -f framehash -
stream_index, packet_dts, packet_pts, packet_duration, packet_size, hash
0,          0,          0,        1,   460800, 29bcc2db3726c7dfec1826c5740f603f
0,          1,          1,        1,   460800, b5fdc23d93cbd043dc2b9290dc8378f0
0,          2,          2,        1,   460800, ee0709942f24b458fd2380d134dcb59d
...

$ ffmpeg -i output.mkv -map 0:v -f framehash -
stream_index, packet_dts, packet_pts, packet_duration, packet_size, hash
0,          0,          0,        1,   460800, 29bcc2db3726c7dfec1826c5740f603f
0,          1,          1,        1,   460800, b5fdc23d93cbd043dc2b9290dc8378f0
0,          2,          2,        1,   460800, ee0709942f24b458fd2380d134dcb59d
...

In the above examples each associated frame for the input and output share the same hash ensuring that the frames are identical and that the output is lossless.

Also see

• FFmpeg image file demuxer documentation
• FFmpeg Wiki: How do I encode single pictures into movies?

Answer 2

This will output a lossless H.264 video where frames will use information from other frames

ffmpeg -f image2 -r 30 -i %09d.jpg -vcodec libx264 -profile:v high444 -refs 16 -crf 0 -preset ultrafast a.mp4

Explanation of options:

• -f image2 - tells ffmpeg to select a group of images
• -r 30 - tells ffmpeg to encode at 30 frames (or images) per second (change this to your desired framerate)
• -i %09d.jpg - tells ffmpeg to use the images 000000000.jpg to 999999999.jpg as the input. Change the 9 in %09d.jpg to how many zeroes the names of your image sequence has. If your file names are, for example, img0001.jpg, then this would be expressed as img%04d.jpg
• -vcodec libx264 - tells ffmpeg to output to a H.264 compliant file
• -profile:v high444 - tells libx264 to use the High 4:4:4 Predictive Lossless profile, allowing lossless encoding
• -refs 16 - tells libx264 to have 16 images stored in a buffer, so that they may be referenced by other images in the video
• -crf 0 - tells libx264 to perform a lossless encode
• -preset ultrafast - tells libx264 to prioritise encoding speed over output file size
• a.mp4 - tells ffmpeg to save the output in an MP4 file called a.mp4. Change this to the file name and format that you wish to use

Answer 3

You can create an avi animation as a series of png images ( png is lossless so the jpeg => png conversion should not degrade your pictures ):

if your images a named img_0001.jpg

ffmpeg -r 25 -start_number 1 -f image2 -i "img_%04d.jpg" -vcodec png video.avi

where "25" is the frame rate you want in the resulting video. -start_number is not needed if it is 1, but it is useful if your first video number is not 1.

If you want to encode in mjpeg with highest quality command line is:

ffmpeg -r 25 -start_number 1 -f image2 -i "img_%04d.jpg" -vcodec mjpeg -qscale 1 video.avi

And the beauty of the thing is that you can convert the video back to a series a pictures:

ffmpeg -i video.avi "img_series_%04d.png"
ffmpeg -i video.avi "img_series_%04d.jpg"

etc...

Answer 4

To expand on LordNeckbeard's answer, yes, just mux the JPEG data into an MJPEG video stream. That will be the smallest representation of the exact sequence of output images, even though MJPEG is a terribly inefficient codec by today's standards. (no temporal redundancy, and not even any intra prediction.

You can make a variable-framerate MJPEG video to take advantage of the duplicate images in your input.

ffmpeg -framerate 30 -i input%03d.jpg -vf mpdecimate -codec copy output.mkv  # doesn't work.

Hrm, this isn't going to work, since mpdecimate won't work on compressed data, and we can't let ffmpeg decode and then re-jpeg the image data without loss and CPU cost.

Maybe if you replaced duplicate jpg source files with empty files with that sequence number, or something?

Since this question isn't even recent, I'm not going to take the time to figure out how to do it unless someone replies to ask how. But since MJPEG can go into an mkv container, I'm sure it's possible to have a file that doesn't duplicate the jpeg data for repeated frames, but instead just doesn't have an output frame to decode until the sequence of duplicates is over.

Oh here's an idea:

ffmpeg -framerate blah -input blah -vf mpdecimate -f mkvtimestamp_v2 mpdecimate.timestamps

Then remove (or move aside) all the jpegs for frames that mpdecimate wants to drop (probably it has some logging options? Or -vf showinfo, and parse that, and move or hardlink only the frames that show up in its output, leaving behind the dropped JPEGs?). mux that to a MJPEG.mkv, then do something with mkvmerge to replace the frame timestamps in that with the timestamps from mpdecimate.timestamps.

If you were xcoding, instead of just muxing jpeg data to MJPEG, this would be MUCH easier, since you'd just use my first command with mpdecimate and any codec other than copy, and it would Just Work(tm).

I haven't tried any of this, since this was an old question. Also the reason I haven't filled in the gaps of how to actually filter your directory of jpegs based on mpdecimate output, or how to actually use the timestamp stream.