Deep Angel - AI that erases objects from images

@rrhoover Oh just you wait! That doesn't yet exist, but we have something in the pipeline very similar to that concept. As a grad school project at the Media Lab, Deep Angel is a speculative design piece and a great springboard for inspiring new ideas and figuring out something really useful.

Thanks for the question @ems_hodge! The philosophy side is intentionally a bit hidden on the website. If you check out the concept page (http://deepangel.media.mit.edu/a...), you can get an initial inkling of the philosophy behind what we’re doing. And if you’re into mysteries, there’s quite a few Easter eggs and a rabbit hole to discover. Unlike most applications that serve content, we’re providing an opportunity for Internet people to remove the objects that mediate our world and social relations. One might consider this to have a Situationist flare. We want people to explore. In many ways, the Internet has turned inwards in the sense that users more often act upon calculated design choices and algorithmic recommendations rather than an innate curiosity. Here’s an attempt (and admittedly, just a start) to encourage discovery and explore how absence can make us think. In Zen Buddhism, the concept of emptiness represents not void nor nothingness but rather the non-limitations and nondefinition of the infinite. It’s hard to explain, and we’re trying to communicate this concept through the experience of Deep Angel.

@kaigradert 100% agree on how time consuming it is currently!! The concept behind Adobe Cloak and Deep Angel is the same. With Adobe Cloak, you manually specify the region that you want to remove whereas Deep Angel does this based on a text input (e.g. remove a person or stop sign or street lamp). Once the object is removed, then Adobe Cloak and Deep Angel are conceptually the same. They are both "inpainting," filling in the image with what makes sense from context. I believe Adobe Cloak uses the Patch Match under the hood whereas Deep Angel uses a deep learning architecture. We're still a bit away from commercial viability, but it's getting closer.

Thanks @kcucchia!! That was Zivvy Epstein, Max (the dog), and me with a lot of help from our friends!

@kcucchia the worst video I have ever seen, usually I never skip parts of promo videos lol

@baybayin Thanks!! It's fascinating what happens when the AI doesn't have enough context to re-create a plausible background. We've been calling it "mean collapse" since what's happening is the pixels are collapsing into the average color of the surrounding image. This particularly happens when the AI hasn't been trained on similar backgrounds before. You'll notice darker photos, photos with objects on the edge, and photos with very large objects often get that ghostly edit. Would love to collaborate with you on an art project, and I'd be happy to expose the gif API to you if that's useful.

@kaayotee Thanks for the support!

@lewis502 thanks for the feedback! Snapheal is very cool, but we're not looking to compete with it. One of the very cool, a bit scary, and extremely relevant aspects of Deep Angel is how automated the photo manipulation is. There are only two user inputs (1) the photo (2) the object name. Essentially, this reveals technology that already exists can very much manipulate media at massive scale.

@babken_karapetyan Thanks for the support!!