HyperDreamBooth is a hypernetwork designed to efficiently generate personalized models from a single image of a person, overcoming the challenges of time and memory requirements in the process of personalization. The project seeks to tackle the problems of model size and speed in text-to-image personalization, particularly in DreamBooth models. By using HyperDreamBooth, a person’s face can be generated in various contexts and styles swiftly while preserving high fidelity to their identities. The developed method outperforms DreamBooth in speed and efficiency, achieving personalization 25 times faster and yielding a model that is 10,000 times smaller.

 

Publication date: 13 Jul 2023
Project Page: https://hyperdreambooth.github.io
Paper: https://arxiv.org/pdf/2307.06949.pdf