✦
Summary
DreamBooth is a research method for fine-tuning text-to-image diffusion models for subject-driven generation. Given just a few images of a subject, typically three to five, it fine-tunes a pretrained text-to-image model so it learns to bind a unique identifier to that specific subject, then synthesizes fully novel photorealistic images of the subject in different scenes, poses, views and lighting conditions.
Key Features
Fine-tunes text-to-image diffusion models from just 3-5 photos of a subject
Binds a unique identifier to the subject for prompt-based generation
Subject recontextualization in new scenes, poses, views and lighting
Text-guided view synthesis from specified viewpoints
Appearance and property modification while preserving identity
Artistic rendering of subjects in the style of famous painters
Class-specific prior preservation loss to retain the model's general knowledge
Open research project with published paper and dataset
About DreamBooth
The technique uses a class-specific prior preservation loss that leverages the model's semantic prior, so the fine-tuned model retains its general knowledge while learning the new subject. It enables previously difficult tasks including subject recontextualization, text-guided view synthesis, appearance and property modification, artistic rendering in the style of famous painters, and accessorization, all while preserving the subject's key features.
DreamBooth is aimed at researchers, artists, designers, educators and content creators interested in personalized image generation. It is published as an open research project with a paper and dataset, and the method works with models such as Imagen and Stable Diffusion rather than being a commercial product.