Why
Dreambooth for Stable Diffusion dropped and suddenly you could train a model on your own face. Everyone started making apps for this. Everyone made them for the wrong platform.
So I made one for iOS.
How
This was a build-in-public project. I wrote down every step. A few things I refused to compromise on:
- Minimal friction from open to payment
- As cheap as possible to operate, so I rolled my own API
- Every format Photos will display — RAW, HEIC, JPEG, PNG, etc
- Encourage people to come back
The UX
A few things put it ahead of most AI avatar apps (besides being native SwiftUI): Sign in with Apple, iCloud sync, on-device upscaling — the user gets what they want and I don't pay for GPUs — bad-sample detection (missing faces, sunglasses, anything that tanks output), and more.
Even a custom notification sound.
The Quality
Pretty early in the Stable Diffusion experiments I figured out a few things that made the output actually look like the person, with a more proportional, realistic face. A lot of AI photo apps still struggle with that.
Makes me proud.
Conclusion
First serious dive into ✨ AI ✨ machine learning, and there will probably be more. Follow me on the X-dot-com world wide web site application for whatever I try next.