You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
I don't think I've made my intentions clear on what our work on the restoration model is trying to achieve, so I'll explain. What we're aiming to do have something that takes any photo that you throw at it, and fixes whatever is "wrong" with it. 1/https://twitter.com/citnaj/status/1316636605558394881 …
What's the definition of "wrong"? In our case, deviations from what would typically be expected of a good modern photograph. That can be noise, color fading/distortion, over-exposure, being a bit out of focus (blur), or even jpeg artifacts. 2/
The key concept though is that we want this to look natural. Typically when you see deep learning models modify images, there's "tells" that it's not very natural. We're trying to avoid that as much as possible. So "just enough" intervention is the goal. /3
Thus, you won't see dramatically enhanced results like you do with something like Remini- where faces are magically derived from very little image evidence. And to be clear- that's quite an amazing feat. But it comes at a cost and that is that it has a distinct "look". /4
i.e. we're trying to make it so you just can't tell if a model was used or not. We're calling it the Anti-Photoshop. Rather than buying/learning Photoshop, you just run it through a single model in less than a second and it does pretty much what you wanted it to do. /5
Fascinating! is there a way to support/contribute?
Well our current source of funding is MyHeritage, so paying them means supporting us financially. We'll be continuing that path for the foreseeable future because we have absolute freedom to pursue research that we think should be done. 1/
Awesome! Have you achieved a GAN-free model(s) for this task? I remember you mentioning that U-Net-like architectures and attention (not GANs) were giving you great results for DeOldify colorization. Are these great results coming from a similar architecture?
Exact same architecture/training actually. The grand goal has been to find a good general approach to image to image.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
