4 ms·

>ML has allowed computers to generate realistic images that humans cannot discern I've yet to see a deep fake that's even remotely indiscernible from the real

>ML has allowed computers to generate realistic images that humans cannot discern I've yet to see a deep fake that's even remotely indiscernible from the real thing.

andrewnicolalde 7y ago
How do you know?
- EpicEng 7y ago
  Well, technically I don't, for the reason I don't know if the abominable snowman exists. I have yet to see one I knew to be a df which was even close to being convincing.
  - joycian 7y ago
    Have you seen CTRL Shift Face, the YouTube channel?
afterburner 7y ago
You probably haven't been noticing them in movies though. Such as when they stuck Margot Robbie's entire head on an actual skater for I, Tonya. https://www.youtube.com/watch?v=bqgD6lHrQ_8 https://www.youtube.com/watch?v=bqgD6lHrQ_8
- EpicEng 7y ago
  Was that a DF though?
  - BubRoss 7y ago
    That video shows a CG head and says CG face replacement. Typically sequences like this are done on a shot by shot basis and end up being a very healthy mix of compositing, cg face replacements, CG head replacements, etc. These types of effects have been used for literally decades across a wide range of movies and have nothing to do with the recent trend and techniques of 'deep fakes'.
    - EpicEng 7y ago
      Right, so not a DF, which is what we're talking about here.
      - BubRoss 7y ago
        I was answering your question.
  - JaRail 7y ago
    Gemini man is a better example of where this is heading. https://www.youtube.com/watch?v=k5y4kxhZIBA https://www.youtube.com/watch?v=k5y4kxhZIBA A DF has a pretty specific meaning if you are diving into the technical aspects. It's an approach that has a lot of specific limitations that put a practical ceiling on how good it can be. For example, it's a 2D system that won't capture lighting differences with good accuracy. The limitations become more apparent as resolution increases. Newer systems use source material to generate 3D facial models. The face in the target video is also 'motion captured.' The target video's lighting is estimated and a resulting face image is rendered and composited into the target. That's a lot closer to the pipeline of movies such as Gemini man. So while DF isn't a technique used in movies, machine learning is heavily involved. It's built into model generation, motion capture, animation, etc. Artists will start with captured/generated assets and hand-tweak them to get the desired result.
blackearl 7y ago
All it takes is a little filtering, some camera shake, and you can hide the tells of a fake.

>ML has allowed computers to generate realistic images that humans cannot discern I've yet to see a deep fake that's even remotely indiscernible from the real thing.

andrewnicolalde 7y ago

How do you know?

EpicEng 7y ago
Well, technically I don't, for the reason I don't know if the abominable snowman exists. I have yet to see one I knew to be a df which was even close to being convincing.
- joycian 7y ago
  Have you seen CTRL Shift Face, the YouTube channel?

afterburner 7y ago

You probably haven't been noticing them in movies though. Such as when they stuck Margot Robbie's entire head on an actual skater for I, Tonya. https://www.youtube.com/watch?v=bqgD6lHrQ_8 https://www.youtube.com/watch?v=bqgD6lHrQ_8

EpicEng 7y ago
Was that a DF though?
- BubRoss 7y ago
  That video shows a CG head and says CG face replacement. Typically sequences like this are done on a shot by shot basis and end up being a very healthy mix of compositing, cg face replacements, CG head replacements, etc. These types of effects have been used for literally decades across a wide range of movies and have nothing to do with the recent trend and techniques of 'deep fakes'.
  - EpicEng 7y ago
    Right, so not a DF, which is what we're talking about here.
    - BubRoss 7y ago
      I was answering your question.
- JaRail 7y ago
  Gemini man is a better example of where this is heading. https://www.youtube.com/watch?v=k5y4kxhZIBA https://www.youtube.com/watch?v=k5y4kxhZIBA A DF has a pretty specific meaning if you are diving into the technical aspects. It's an approach that has a lot of specific limitations that put a practical ceiling on how good it can be. For example, it's a 2D system that won't capture lighting differences with good accuracy. The limitations become more apparent as resolution increases. Newer systems use source material to generate 3D facial models. The face in the target video is also 'motion captured.' The target video's lighting is estimated and a resulting face image is rendered and composited into the target. That's a lot closer to the pipeline of movies such as Gemini man. So while DF isn't a technique used in movies, machine learning is heavily involved. It's built into model generation, motion capture, animation, etc. Artists will start with captured/generated assets and hand-tweak them to get the desired result.

blackearl 7y ago

All it takes is a little filtering, some camera shake, and you can hide the tells of a fake.