We worked with Google to help prototype the experience and shape the app’s integration of Gemini, Imagen, and ML Kit. The result? An open-source demo that shows how structured AI outputs, multimodal prompts, and on-device ML come together to create avatars that feel personal, expressive, and just plain fun.
Our role focused on finetuning the Imagen model to bring these bots to life.
We led the effort to fine-tune the Imagen 3 model, turning a general-purpose text-to-image model into something far more playful, characterful, and Androidify-specific.
Using LoRA (Low-Rank Adaptation), we trained the model on a wide range of Android bot assets with varying colors, accessories, and styles to dial up the charm and variety of outputs. We generated custom image–text pairs to create a rich training dataset and ran a full fine-tuning pipeline that balanced efficiency with expressiveness. The result is a model that better understands the visual language of Androidify, from the subtleties of sunglasses and hairstyles to the joyful weirdness of bot proportions.
The demo app currently uses the baseline Imagen model, but the fine-tuned version—along with a custom Firebase AI Logic SDK—is set to launch later this year. The difference is not just aesthetic: it’s about making generative AI feel more intentional, delightful, and brand-aligned.
