DreamBooth is a technique to fine-tune the entire Stable Diffusion model on a small, personalised dataset focused on a particular subject, like your face, your pet, or any unique object.

How does it work?

• You prepare around 10-20 images of the subject you want the model to learn.
• DreamBooth fine-tunes the full model weights (not just embeddings) to deeply understand this subject.
• The training uses class-specific regularisation to prevent the model from forgetting other knowledge.
• After training, you can generate images of that subject in many different contexts and styles, with high fidelity.

Why use DreamBooth?

• Produces very accurate and detailed results.
• Great for highly personalised or niche concepts where detail matters a lot.
• You can put the subject in different scenes, styles, or compositions.

Tradeoffs:

• Requires much more VRAM and longer training time (hours vs minutes).
• The model size and complexity make it less flexible to share or combine compared to LoRA.
• Fine-tuning risks overfitting if not done carefully.