michaelni changed the topic of #ffmpeg-devel to: Welcome to the FFmpeg development channel | Questions about using FFmpeg or developing with libav* libs should be asked in #ffmpeg | This channel is publicly logged | FFmpeg 7.1 has been released! | Please read ffmpeg.org/developer.html#Code-of-conduct
iive has quit [Quit: They came for me...]
arch1t3cht6 has joined #ffmpeg-devel
arch1t3cht has quit [Ping timeout: 265 seconds]
arch1t3cht6 is now known as arch1t3cht
Marth64 has quit [Quit: Leaving]
thilo has quit [Ping timeout: 272 seconds]
thilo has joined #ffmpeg-devel
jamrial has quit []
System_Error has quit [Ping timeout: 264 seconds]
<aaabbb>
was this a dream or did i read something on the mailing list a while ago about the possibility of a lossy feature in a future ffv1 for use as an intermediate codec?
System_Error has joined #ffmpeg-devel
Martchus has joined #ffmpeg-devel
Martchus_ has quit [Ping timeout: 260 seconds]
^Neo has quit [Ping timeout: 248 seconds]
<Lynne>
I have a very faint memory of that, but as someone who's written a complete encode and decode implementation, its possible, but I wouldn't advise it
<Lynne>
the range coder is slow and rather hacky
<aaabbb>
but fast enough to encode and decode in real time
<Lynne>
yeah, no, it wouldnl't be fast enough even if you cut down the number of bits coded by half
<Lynne>
in software, anyway
<Lynne>
for mezzanine codecs, you really want some frequency decomposition so you can play around with the spectrum and push the error out in less visible parts
<aaabbb>
it encodes at 1.5x on my laptop for 720p
<aaabbb>
and only a little under 1 at 1080p and my laptop sucks
<Lynne>
not fast enough for the 4k era today
<aaabbb>
hmm
<aaabbb>
it's just that there are no mezzanine codecs in ffmpeg that are any good i thought. like the prores encoder is really basic and nothing compared to apple's native
<nevcairiel>
apple will always whine about any software they dont fully control
<JEEB>
also that by now is a quite an old page
<JEEB>
plus they go as far as saying "might", instead of "will"
<JEEB>
(this is the "licensed and unlicensed prores implementations" page)
galad has quit [Remote host closed the connection]
galad has joined #ffmpeg-devel
Guest84 has joined #ffmpeg-devel
<Lynne>
aaabbb: vc2_vulkan is quite good, and will be merged soon I hope
<Lynne>
its exactly what you want - very fast, can be lossless, its got better quality than prores if you give it more bits because wavelets are inherently lapped and need no filter
mkver has joined #ffmpeg-devel
<aaabbb>
how does it compare to prores at the same bitrate?
<Lynne>
depends on the slice size
<Lynne>
in general it ought to be better at the typical slice sizes of 32x32 and up
<aaabbb>
oh cool as long as it's not significantly worse
<aaabbb>
if it's a mezzanine then obviously i'm not looking for stunningly good compression for a given bitrate :)
<JEEB>
I recall like 10-15 years ago people tested H.264 as an intermediate format and GOP up to 3 frames worked fine
<JEEB>
if I had to have a fast decode'able format for intermediates, I'd probably just utilize a low enough CRF :P esp. if the editor has something like mainconcept's decoder licensed, which means you get 10bit and 4:2:2 and all.
<Lynne>
panasonic/sony's professional cinema cameras still use that from of h264 422 as the main raw format
<Lynne>
its awful, why would you ever buy one when you can get a blackmagic's cameras which give you true raw bayer with mild compression
<JEEB>
they also now reuse that naming for their HEVC thing
<fflogger>
[newticket] ffvii: Ticket #11378 ([ffmpeg] ffmpeg can't detect nested m3u8 for subtitles) created https://trac.ffmpeg.org/ticket/11378
<JEEB>
also the all intra mode was high enough bit rate at least in the recent stuff for sony. although yes, if you can take full res/no subsampling then capturing through the capture interface on cameras is usually better
<JEEB>
I do wonder how subsampled the HDMI raw data output is in f.ex. panasonic GH6, where they advertise cinema 4K at 120/1.001fps and 12bit.
<jamrial>
that sample also has a lot of metadata obus of type 6, which is one of the "user unregistered", and libdav1d prints a lot of warnings about it
<jamrial>
maye it should not for types 6 to 31
rvalue has quit [Read error: Connection reset by peer]
rvalue has joined #ffmpeg-devel
Traneptora has joined #ffmpeg-devel
<Traneptora>
Masaru Nomiya <nomiya@lake.dti.ne.jp> seems to be giving AI answers on ffmpeg-user email list
<Traneptora>
they have done it twice now, can we block them?
<jamrial>
ai, or machine translation?
<Traneptora>
jamrial: looks AI generated. starts with a preamble and gives a nonsensical answer
<Traneptora>
check the answer about non-monotonic DTS. also check the answer about pkg-config libass (the person clearly didn't look at the log file that was posted)
<ePirat>
Traneptora, do you hav a link to the pkg-config one?
<ePirat>
Traneptora, yeah this does look very much AI generated…
<Traneptora>
ye, maybe instead of blocking them give them a warning that if they keep doing it they'll be banned from the list
<Marth64>
wow good eye
<BtbN>
One thing that confuses me a bit right now in the ML discussion is that it's somewhat painted like michael is blocking a migration to Forgejo/Gitlab. Which from my perspective is not the case at all.
<BtbN>
It was one of the first things I discussed with him on how to practically implement when I got root access, and I feel like we got to a good result, that kinda just needs implemented now.
<jamrial>
then i'm even more confused than i thought
<j-b>
Could we just stop and have a break for a few days?
<j-b>
This is more than tiring.
<BtbN>
Talking about "AI" nonsense. I just watched a video about speedrun cheating the other day, where someone asked ChatGPT on how to identify a cheater by audio spectrum analysis.
<BtbN>
A short time later, he falsely accused someone of cheating
<Traneptora>
Marth64: wasn't my eye, one of the first responses to the DTS question was "please don't use AI answers"
<j-b>
BtbN: AI is amazing on things. And catastrophic on others.
<j-b>
BtbN: but people don't know how to use it.
<BtbN>
What was crazy about it was that first he just directly asked ChatGPT "Yo, is this dude cheating? Here's a picture of the audio spectrum."
<BtbN>
And ChatGPT confidently explained why he is clearly cheating. But apparently even the dude doing that realized THAT is nonsense, given it said that about every sample he gave it.
<j-b>
BtbN: have you asked ChatGPT if you were a minister from your country ?
<BtbN>
I have so far only used ChatGPT as Google
<BtbN>
And it has failed most of the time
<j-b>
because it's useless for that.
<BtbN>
And failed in such a way that it wasted a lot of my time, to figure out if what it just said makes any sense at all
<BtbN>
Would have been faster to just research myself
<JEEB>
with the bing edition that includes interwebs, it actually once found a github repo of interest for me
<j-b>
I use it quite often, and it is amazing for some tasks.
<BtbN>
I asked it how to fix Japanese YT subtitles to use Chinese characters instead of Japanese
<BtbN>
It gave me 6 answers. All wrong.
<BtbN>
Took me an hour to come to that conclusion
<BtbN>
correct answer was not hard to find in google and quick to do
<ePirat>
jamrial, can I somehow convert a mp4 with iamf to a "pure" .iamf file?
<JEEB>
font locales are fun
<JEEB>
or well, font rendering based on it
<jamrial>
ePirat: yeah, just remux it
<BtbN>
What I don't understand is... WHY are those clearly different characters the same Unicode codepoint?!
<ePirat>
jamrial, doesnt work
<ePirat>
jamrial, I just get [iamf @ 0x617000012d80] There must be at least two stream groups
<ePirat>
[out#0/iamf @ 0x611000007700] Could not write header (incorrect codec parameters ?): Invalid argument
<JEEB>
BtbN: back in ye olden days there was this whole thing about han unification...
<jamrial>
ePirat: it's not automatic. you need to either define stream groups manually, or map the ones from the input for the output
<BtbN>
直 is just a completely different character, depending on the font in use
<jamrial>
see the fate tests
<BtbN>
If it has an "L" shape in the bottom/left, it's Japanese. If not, it's Chinese.
<JEEB>
so if the meaning/symbol behind it was matching, they considered it a rendering thing to figure out which locale's edition to utilize