ffmpeg-concat
ffmpeg-concat copied to clipboard
transitive-bullshit
optimize temp frame creation and transcoding
This PR implements the approach discussed in #8, namely it makes a tradeoff between compute time versus temp storage requirements that should make the program more robust for larger input videos.
Previously, we first extracted all frames from each input video, then rendered the transitions where appropriate, then did a final transcoding pass that was very efficient since it only had one input source as a single stream of frames.
Now, we only extract frames from input videos that are actually used in the transitions, which for larger input videos can be a significant reduction in temp frames extracted. This has the disadvantage, however, that the final transcoding phase is more complex and involves O(n) inputs instead of O(1) inputs.
If we decide to merge this approach, it will necessitate a major version bump to v2.0.0.
I would like to gather some performance numbers for speed and memory usage before committing to this approach. In my very preliminary tests, the speed has remained approximately the same with a significant reduction in temporary disk and memory usage, though I'm guessing that we should expect speed improvements from longer / larger input videos.
@BrandonCookeDev, @daniel-habib, I'm guessing these changes will interest you. Thoughts & feedback are more than welcome.
@mergebandit this would seriously reduce some of the issues we've seen with prod Automagical renders running out of disk space in the past.
@XavierTS The PR works fine for 95% of cases, but I've run into an issue with a more involved test case where ffmpeg just halts about 70% of the way through and haven't had time to follow-up.
@transitive-bullshit : what's preventing you from merging feature/optimize-frames into master? in other words: how can i help? just run a ton of use cases? or do you already have specific use-cases that cause issues?
The only problem with this approach is that the ffmpeg inputs become more complicated instead of a single stream of frames.
After running a startup focused around video creation for several years, one hard & fast rule I've found when dealing with ffmpeg is that the simpler your can make your inputs, the better your usage will scale in terms of video sizes and different types of inputs.
I have a few test cases for my previous company that this approach will randomly stall on, and aside from it likely being a bug somewhere in ffmpeg, there's not much I can do to resolve the issues.
One possibility would be if I published a fork of ffmpeg-concat that took this fundamentally different approach, as it will still likely be useful for many people.
Thoughts?
ffmpeg and keeping-it-simple - agree. you never know if some rare option was rigorously tested or even designed to support all uses cases.
having 2 active forks sounds bad. maintenance-wise: every fix has to be applied to both, ux-wize: developers would not know which of the forks is better.
i can take this branch, build my app based on that, and run some tests on it for a while if that would help with confidence in ffmpeg current options. would that help?
That'd be great @daniel-habib. I'd expect this branch to work for 95% of use cases.
nothing from my side, sorry. priorities changed, and i did not have resources to test this :( apologies for not notifying before
With the version on master, I was not able to take a 10 minute video and concat an intro and an outro to it (244 KiB Intro, 42 MiB talk, 428 KiB Outro). More than 30 GiB of /tmp space would have been required. I can't imagine how much space the one-hour talks would need.
By using this version:
$ nvm use 8
$ npm install -g https://github.com/transitive-bullshit/ffmpeg-concat.git#feature/optimize-frames
It works like a charm, thanks!
Great to hear @soerface. You can also use the --frame-format png option to store each frame compressed instead of the faster uncompressed raw format that's the default.
Is this branch still actively developed?
The package.json must be updated!
Swap it with the master branch one.
This works pretty well for my usage, however for some reason the output does not contain audio.
Since I edit files with 3-30 GB and only want to add an introduction of 10 seconds, this would make it possible to automate this process. Unfortunately its unfeasible to use the version on master even with --frame-format png
@badosu : not having audio on a result is kind of a "feature' i have tried to address it in https://github.com/transitive-bullshit/ffmpeg-concat/pull/46 which was merged to master once. one could try to merge master into this and see where it takes us. looks like it is something that needs to be done anyway to resolve conflicts. i would have done that but: 1/ i am not sure why this was not approved in the first place. if there's a good reason for it and we don't address it - we'll just create a bad master. 2/ some people where not pleased with my "fix". i think it was cases of bad input, but still... we might want to address that before merging here. in the meantime: consider running another ffmpeg to concat the into a big audio after you have a big video. and then ffmpeg to merge the 2.
Well, for often used cases the current functionality might be enough but I would expect to be able to concatenate videos with audio without scaling space usage as much as it is at the moment (with resolution x duration).
If I have to make additional commands it loses a bit of purpose (which I assume to be user friendly to ffmpeg concatenation commands).