Two years of 3D faces on a phone: what I got wrong
Leaving the selfie-to-3D-face app I helped build out of my thesis. The technology mostly worked. The rest is the interesting part.
This is my last week on the 3D face app. I joined two years ago to turn my thesis into a product: one selfie in, an animated 3D version of your face out, running on the phone. I’m moving full time to the app studio I’ve been working with, and before the details fade I want to write down what I’d do differently.
The technology mostly worked, which is not something I expected to be able to say in 2013. We got landmark detection and mesh fitting running at interactive speed on an iPhone, we fixed the dead-eyed look with a lot of small motion tricks, and the output looked good enough that people’s first reaction was surprise and not horror. For a small team on phone hardware, I’m proud of that.
I got three things wrong, and they’re connected.
I thought quality would sell the product. Every month I’d find something that looked wrong and fix it, and every month the faces got better, and usage didn’t follow. People who tried the app liked it. The problem was that not many people tried it, and the ones who did usually didn’t have a reason to come back after the first day. A better face didn’t change either of those things.
I treated the 3D face as the product. It was really a feature looking for a context. Put yourself in a game, in a video message, as a sticker in a chat. Those are products. A face that you can rotate and make smile is a demo, and people treat demos as demos: impressive once, then closed. We talked about integrations and messaging use cases, but we kept prioritizing the engine because the engine was what we knew how to improve.
I underestimated distribution. We launched into an App Store with over a million apps and expected good reviews to carry us. The apps that grew around us had a built-in way to spread, like sharing something funny made in the app, or an existing big audience. We had neither, and I thought of that as a marketing problem for later when it was really a product design problem from the start.
I’d also say something about timing. In 2013 face tracking on phones was a research-grade problem. By now, a year and a half later, apps like Looksery already do real-time face effects on phones, and Facebook’s DeepFace shows what large companies with lots of data can do. The capability we worked hard for is becoming something platforms will provide. When that happens, being early on the technology doesn’t help much unless you’ve already built the audience.
What I’m taking with me: I know how to take research to shipped software on constrained hardware, and I know much better now that shipped software isn’t the goal. At the studio I’ll be building apps for clients across a lot of industries, which should be a fast way to learn what people actually use. I’ll probably write more about faces anyway. I don’t think I’m done with them.