Shotcut 24.10 Open-Source Video Editor Adds Initial AI Feature

[caligula]

AFAIK it's much faster than kdenlive, since it supports GPU effects. The maintainer of shotcut is also the maintainer of MLT.

Kdenlive is pretty slow considering it's written in C++. Just opening and closing a project takes ages. You can see from the debug messages in the log that something is totally fucked up there.

[ehansin]

Olive video editor...

Thanks for the heads up and some details on what you like about it. I have heard of it and when reminded here remembered the 0.2 rewrite stuff and talk of it being possibly a really good option with some unique features. I just installed via Brew on the macOS laptop I'm on right now. Looks like a pretty old version (says April 2019!), but will at least try and I see the Olive site has a version for 2024, so maybe that is a better option

[pong]

Caution. While the Whisper code and "model weights" are lic. MIT, I don't see any way of auditing the source material those language models were trained on. This could still be a legal landmine given the state copyright is in today.

Very abstractly I see your point about ML models "in general". But at the coarsest "pragmatic" level of risk wrt. the end user my initial reaction is to think that it would "matter" more to the end user when using generative ML models and creating generated content that they expect to be free of IP infringements but might have some if the ML model regurgitates nearly literally substantial parts of its (possibly non open to for non transformative reuse) training material as the output.

So in that case, generating images, generating code, generating music might be more areas of concern.

But whisper basically AFAIK generates more or less individual tokens and those are more or less syllables or something like individual words or coarse descriptions of non-speech sounds etc. It's not AFAIK likely to generate absent 95% literally corresponding input the poetry of shakespeare or the lyrics of an opera song or anything else that's of sufficient length / uniqueness / etc. which one would likely be subject to copyright / IP.

At worst in any ordinary circumstance that comes to mind I suppose it MIGHT be able to generate text that describes something using some trademarked descriptive text like "Growl like the Hulk" based on some input sound, but what you're in theory
going to get out is a ~90% accurate "transcription" of your input audio word by word so if you've got output IP concerns, you
probably have the same with the input.

The more abstract case would be merely having / using the model regardless of its output could be something some troll could
complain about wrt. patent or made-using-my-IP infringement claim affecting many users or users with use case workflows someone is trolling but that's very different and in many cases transcends having to worry about whether the output is "clean/licit".

[sophisticles]

Caution. While the Whisper code and "model weights" are lic. MIT, I don't see any way of auditing the source material those language models were trained on. This could still be a legal landmine given the state copyright is in today.

I'm not too worried about that, I assume that they have done their due diligence.

[Quackdoc]

Thanks for the heads up and some details on what you like about it. I have heard of it and when reminded here remembered the 0.2 rewrite stuff and talk of it being possibly a really good option with some unique features. I just installed via Brew on the macOS laptop I'm on right now. Looks like a pretty old version (says April 2019!), but will at least try and I see the Olive site has a version for 2024, so maybe that is a better option

yeah that would be the 0.1 version, 0.2 has a lot of changes, but some folk are still on 0.1 for various reasons