After we built the VTO at Warby, we realised that it was actually somewhat hard to use if you wore glasses, because you would need to take your glasses off and then is difficult to see what you look like with the glasses on.
The obvious answer was to either record some video or take some photos and then show them back to the user when they’ve put their glasses on.
But one random hackathon I thought it would be cool to see if we could train a model to take a image of someone with glasses on then remove their glasses (and then we could place virtual glasses on in a second pass after that)
This was all long ago, before there were diffusion based models, and we were futzing around with pix2pix and GANs 👴🏻👴🏻👴🏻.
