Since there's an onnx model this should be possible to export to onnxruntime for web, i.e. a real one click (no installation) deepfake straight on the web.
The readme sounds angry. When you run this without ffmpeg, it prints:
>python run.py
ffmpeg is not installed. Read the docs you idiot.
ffmpeg is not installed. Read the docs you idiot.
ffmpeg is not installed. Read the docs you idiot.
ffmpeg is not installed. Read the docs you idiot.
ffmpeg is not installed. Read the docs you idiot.
ffmpeg is not installed. Read the docs you idiot.
ffmpeg is not installed. Read the docs you idiot.
ffmpeg is not installed. Read the docs you idiot.
ffmpeg is not installed. Read the docs you idiot.
ffmpeg is not installed. Read the docs you idiot.
I found a use for face-swapping recently, when working with Stable Diffusion animations. The main challenge there is keeping the frames consistent between each other. Generating a face and then face-swapping it on the full animation helps with the consistency.
Such software attract all kind of people including the casuals who end up creating github issues for problems that can be solved with a single google search or reading the docs. I got tired of multiple issues regarding ffmpeg and ended up dropping this in there. Not very proud.
Honestly, I like the abrasiveness. It adds a certain charm to the program when it directly chastises you for not reading the provided documentation.
For such situations like this where they *should* be following the setup procedures, the errors should be explicit & straightforward: Follow the setup docs.
This gets really spicy when it can be adapted to run realtime so you can map a face on a Teams call, etc. and combine with voice mapping so you can seamlessly impersonate someone for those "It's me" extortion calls to grandmas. All of those hours of Tiktoks are now an endless goldmine for scammers...
On the other hand, since you mentioned Teams, I foresee some applications in the space of faking presence on meetings on which everyone is required to have their camera turned on.
For this you can already use a prerecorded video as a camera from OBS. The challenge is what to do when asked, and faking someone else doesn't offer any improvement for that.
Actually I think making this kind of AI easier to access is a way to reduce the negative effects of this kind of tools.
If only ML engineers can make deepfakes, people are actually going to believe deepkfakes.
If every 12yo asshole kid flood the internet with deepfakes, then people will actually realize that anybody can fake a video, and stop believing everything they see, which will greatly reduce the potential for misinformation/harassment/the spread of false rumors about people
You could use those things to assist an attack. It's been done many times. But they just don't scale. These types of tools scale. Also I'd love to see someone realtime fake their identity with Photoshop or a pdf editing tool. Seriously send me links I'd love to see it
Aside from entertainment, what's the use of Adobe Premiere? Or Davinci Resolve? Seems odd to exempt the one use case that this is most valuable for. Look at how many Hollywood movies had their CGI improved by fans using this kind of tech and a laptop.
Now imagine what happens when indie film directors have access to the same tech, which simply requires learning these tools.
I'm not dismissing the very real and very legitimate concern about how deepfakes can be (and are being) used for blackmail, extortion, and propaganda. But I wouldn't discount the legitimate uses.
I would be inclined to think this as well. The problem is, large numbers of people will still accept the fakes when it suits their beliefs, and they will claim fake on real footage that doesn't mesh with their beliefs.
This will definitely accelerate the decline of humanity, but we were headed downhill already anyway.
So I assume the author received consent from Scarlett Johansson. Well done, it must be hard to get a hold of her.
[0] https://news.ycombinator.com/from?site=github.com/s0md3v