Dreambooth

DreamBooth fine-tunes diffusion models to generate specific subjects or styles from few example images. Approach: Fine-tune entire model (or LoRA) on images of subject with unique identifier token. Model learns to bind identifier to the concept. Process: 3-5 images of subject → assign unique token ("sks person") → fine-tune model to generate subject when prompted with identifier. Technical details: Fine-tune U-Net and text encoder, use prior preservation (regularization images of class) to prevent language drift, low learning rates. Prior preservation: Generate images of general class ("person") and train on those alongside subject images. Prevents model from forgetting general class. Identifier tokens: Use rare tokens ("sks", "xxy") to avoid overwriting common words. Training requirements: 3-10 images, 400-1600 steps, higher compute than LoRA (full fine-tune), takes 15-60 minutes. Use cases: Personalized portraits, product photography, consistent characters, custom avatars. Limitations: Can overfit, may struggle with very different poses than training, storage for full model weights. Comparison: More thorough than LoRA but less efficient. Often combined with LoRA for best of both.