How to efficiently create a best-palette gif from a video portion straight from the web

Question

I've been using for some time this couple of commands to convert a video segment to an animated gif, having ffmpeg calculate the best palette for it:

ffmpeg -ss $START -i $IN_FILE -t $LENGTH -vf "fps=$FPS,scale=$WIDTH:-1:flags=lanczos,palettegen" palette.png
ffmpeg -ss $START -i $IN_FILE -i palette.png -t $LENGTH -filter_complex "fps=$FPS,scale=$WIDTH:-1:flags=lanczos [x]; [x] [1:v] paletteuse" output.gif

This works very well for local files, but if I start to use remote URLs for $IN_FILE, it downloads the required portion twice - once for the palette generation, once for the actual conversion.

Downloading the full file in advance is generally out of question - often I'm interested in a very small sequence in the middle of a longer video.

I tried to download just the small portion using -ss and -t and saving it - without reencoding - to a temporary file:

ffmpeg -ss $START -i $IN_URL -t $LENGTH -vc copy -ac none temp.mkv

In this case I do avoid the bandwidth waste (only the relevant portion of the file gets downloaded, and only once), but the seek is no longer precise, as -ss on input has only the granularity of key frames for seeking when doing stream copy.

When converting to gif in theory it would be possible to do an extra precise seek and fix this, but there doesn't seem to be a way to get the original timestamp of the start of the temporary file generated above, so it is not possible to calculate to where I should -ss when transcoding to gif. I tried playing with -copy_ts, but it didn't yield anything good.

The trivial solution is to reencode in this first step (possibly applying the scale/resample in the process, to avoid doing it twice later), but I'd like to avoid the cost/quality loss of one extra useless encoding.

So: how can I perform the best-palette video to gif conversion of a small portion of a potentially big networked file, fetching it efficiently (=download once, only the relevant portion), with precise seek and without extra re-encodings?

Answer 1

The problem can be easily solved by using a filtergraph with multiple chains, which allows us to perform the seek/download/filter just once and process it in multiple ways. The seeked/filtered stream is fed to both the palette generator and to the palette application filter, which uses it along with the generated palette. Graphically:

                                       .--> palettegen [pal]---.
 input                                /                        |
 [0:v] -> fps -> scale -> split=2 [a][b]                       V
 with                                 `-> [b] fifo [b] -> [b] [pal] paletteuse -> out.gif
precise
 seek

which translates to:

ffmpeg -ss $START -I $IN_URL -t $LENGTH -filter_complex "fps=$FPS,scale=$WIDTH:-1:flags=lanczos,split=2 [a][b]; [a] palettegen [pal]; [b] fifo [b]; [b] [pal] paletteuse" out.gif

Notice that to use the same stream as input for two separate pipeline branches it's necessary to use the split filter.

Edit: fifo was added thanks to @Gyan's comment; it is necessary because palettegen needs to wait to the end of the stream before generating the palette, and paletteuse cannot start consuming [b] before having the palette, hence if the video is big enough the default buffers for [b] won't be enough and ffmpeg will start dropping frames. The solution is to add a fifo in the middle to handle arbitrary size buffering (care must be taken not to exceed with the video length, as buffering the whole stream in memory can be taxing on the available RAM).

(shameless plug: this is the command I'm now using in my tube2gif_bot Telegram bot)

---

Most importantly, this is a tale about understanding the commands you are using; the second command quoted in the question already used a complex filtergraph, but as I blindly copy-pasted it from the web I didn't really try to understand the opaque filtergraph syntax, so it didn't occur me that just tweaking it a bit would have been the best solution.