DreamBooth

DreamBooth is a technique from Google Research for personalizing text-to-image diffusion models around a specific subject. Given just three to five photos of a person, pet, or object, the method fine-tunes the model while binding the subject to a unique identifier token, after which prompts using that token can place the subject in new scenes, poses, lighting, and artistic styles. The official Google repository does not ship training code; it publishes the evaluation dataset from the paper, 30 subjects across 15 classes covering 9 live subjects and 21 objects, along with the prompts used and per-image license attributions. Practical training implementations live elsewhere, most notably in Hugging Face Diffusers and community trainers such as kohya-ss, where DreamBooth has become a standard recipe alongside LoRA for subject-driven generation. Researchers benchmarking personalization methods and practitioners creating consistent characters or product imagery are its typical users.