DreamBooth is a fine-tuning method for text-to-image diffusion models, published by researchers at Google Research (Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein and Kfir Aberman). A diffusion model creates an image by starting from random noise and removing the noise step by step under the guidance of a text prompt. Fine-tuning means training an already trained model further on new data.
The input is a few photos of one subject, typically 3 to 5, and the name of its class, such as "dog". The method trains the model to link a unique identifier word with that subject, using prompts like "a photo of a [V] dog". A class-specific prior preservation loss keeps the model able to draw other members of the class, so it does not forget what a normal dog looks like. After training, prompts with the identifier place the subject in new scenes, poses, viewpoints, lighting or art styles. The paper used Google's Imagen model, but the method is not limited to it.
DreamBooth is a published research method, not a product. Google's official repository holds the paper's dataset under the CC-BY-4.0 license.