Comment

Comment on Deepfake videos are getting shockingly good | TechCrunch

The example videos are pretty good. Only a few had obvious tells (like the Taylor Swift one), and the rest seemed pretty human-like.

But before everyone goes out and invests in OmniHuman-1 systems, remember that marketing campaigns always show the best they could make, not the average case most people are likely to get. Will it be good enough to trick the average consumer who’s not looking that hard? Maybe. I guess we’ll have to see.

But if all these generative models are so designed to replace the people upon whose videos they’re based, who/what will train the next generation of models, I wonder?

source

Sort:hotnew top

Yingwu@lemmy.dbzer0.com ⁨7⁩ ⁨hours⁩ ago
Until these starts getting used on a broader scale, I’m not convinced these are not schemes to funnel more investment money into their companies. The examples are really short, probably made after I don’t know how many attempts, and probably very limited in what poses and/or actions they can show. I’m so tired of the LLM hype in general.

source
- iopq@lemmy.world ⁨27⁩ ⁨minutes⁩ ago
  People said this about AI generated images like three years ago. Now you have very high quality images generated from a fairly simple prompt. Don’t expect it to stay hard and lower quality forever
  
  source
TheFogan@programming.dev ⁨12⁩ ⁨hours⁩ ago

But if all these generative models are so designed to replace the people upon whose videos they’re based, who/what will train the next generation of models, I wonder?

Maybe we hit a regular cycle…

IE Data is all trained on real video, gets good enough that humans cannot differentiate. Real video becomes rare, AIs are now training on AI videos. Result… AI video becomes effectively copy of copy of copy of copy… degredation becomes obvious. as mistakes are now compounding. AI developers have to start creating and introducing un-tampered video to train with. AI starts to get better.

source