Half a dozen articles on ML-based image manipulation on HN at once. Seems we're really entering into a golden age of AI-based real-world applications, at least in specific niches. Personally I'm really excited about the potential of this in design, art, movies, games and interactive storytelling. Hard to imagine what will be possible in 5-10 years from now, but I kind of expect RPG games with fully AI-generated aesthetics / graphics and stories, where only some core gameplay mechanics are still determined by the designers of the game. Really can't wait to see that.<p>The work described in the linked article is also extremely impressive and feels almost unreal, in any case.
Is it only me who noticed how teeth appears absolutely out of nowhere when people smile on the demo footage? And it looks not facinating. It looks horryfying.<p>Probably because of falling into the uncanny valley [0].<p>0 - <a href="https://en.m.wikipedia.org/wiki/Uncanny_valley" rel="nofollow">https://en.m.wikipedia.org/wiki/Uncanny_valley</a>
You can run it on Replicate here: <a href="https://replicate.com/google-research/frame-interpolation" rel="nofollow">https://replicate.com/google-research/frame-interpolation</a>
This feels like something that would be perfect for one man, or small team, animation studios. If this could draw the in-betweens i imagine a talented artist (which I am not) could produce films in literal fractions of the time it takes to draw every frame. If you're not happy with the result, just add another frame.
I still can't wrap my head around how people absolutely ignore kids' rights to privacy putting their photos/videos without their consent.<p>I would have been pretty bummed by my teens if I found out all my life's history was there for the whole world to crawl, collect, train their ad/surveillance NNs on, etc.
Does anyone know if it's possible to run this on Apple Silicon GPU? I've been playing with Stable Diffusion on M1 and having fun, I'd love to be able to use this to interpolate between frames as shown in another recent post.
> synthesizes multiple intermediate frames from two input images<p>That's a neat use case, and definitely a good way to show off, but what about more than one image?<p>The overwhelming majority of video that exists today is 30fps or lower. The overwhelming majority of displays support 60hz or more.<p>Most high-end TVs do some realtime frame interpolation, but there is only so much an algorithm can do to fill in the blanks. It doesn't take long to see artifacts.<p>I would be more interested to see what an ML-based approach could do with the edge cases of interpolating 30fps video than 2 frames.
It seems like this could be a good way to provide smooth weather / cloud animations using real or raw cloud images rather than those heat maps most apps use.