Lessons from Lensa
In the last two weeks or so, the mobile app Lensa has topped the charts. It's an app with AI-powered photo editing features, but the breakthrough feature has been "AI Avatars" generated with LLMs. For about $4, you get 50 stylized photos of yourself.
While I don't know exactly what they are doing behind the scenes, it's fairly likely they are using fine-tuning Stable Diffusion with Dreambooth. It takes about 20 minutes to generate a set of photos in Lensa, which lines up with the time it usually takes to fine-tune Stable Diffusion with Dreambooth. It should cost about $0.30 for training + 50 inferences on AWS (double that if they are using some higher order infrastructure like HuggingFace).
The Lensa app is published by Prisma Labs, the company behind the 2016 hit mobile app Prisma, which used AI style transfer to produce AI-powered filters in the style of famous artists for photos.
So what did Lensa do right?
Mobile-first is still where most consumers are. If you're doing anything consumer, figuring out how to package it up as a mobile application is the way to go. Not only does it make it easier to reach a larger audience, but it simplifies other parts of the workflow (e.g., no need to upload photos manually, can easily select them through the iOS UI).
Distribution matters. This is an obvious point, but Lensa/Prisma Labs was 7 years in the making. It's easier to launch when you already have a captive audience.
An influencer-focused go-to-market strategy. They tweeted out fine-tuned images of influencers in the space.
No watermark? Maybe this was an oversight, but could have improved virality.