Glossary
DreamBooth
An older full fine-tuning method that bakes a new subject directly into a checkpoint.
DreamBooth (Google Research, 2022) was the first popular technique for teaching a diffusion model a new subject from 3–5 photos. It produces a full new checkpoint (multi-GB) per subject, which is slow, expensive and not stackable.
LoRA has largely replaced DreamBooth in production pipelines because it is 100× smaller and combines cleanly with other LoRAs.