DreamBooth
DreamBooth is a fine-tuning method that teaches a base image model a specific subject - a person, pet, product or style - from just a few reference photos, so it can recreate that exact subject in new scenes.
DreamBooth is a way to personalise an image model. Give it a small set of photos of one subject - say five to twenty pictures of a particular person or product - and it fine-tunes the diffusion model so a unique trigger word reliably renders that exact subject in any setting you prompt.
How it works
Unlike textual inversion, which only learns a new keyword, DreamBooth actually adjusts the model's weights so the subject is baked in. It binds your subject to a rare token and uses a "class" of similar images to stop the model forgetting everything else. The result is the most faithful reproduction of a specific subject - at the cost of a heavy output file, essentially a new checkpoint.
DreamBooth vs LoRA
In practice many people now train a LoRA instead, because it captures most of the fidelity in a far smaller, easier-to-share file. DreamBooth still wins when you need maximum likeness, but LoRA is the lighter everyday choice for a recurring character or style.
Try it in the generator
Put dreambooth to work right now - free daily generations, commercial license included.
Frequently asked questions
What is the difference between DreamBooth and LoRA?
DreamBooth fine-tunes the whole model, producing the most faithful version of a subject but a large checkpoint-sized file. A LoRA trains a small add-on instead, capturing most of the likeness in a few megabytes. LoRA is the lighter, more shareable choice; DreamBooth is for maximum fidelity.
How many images does DreamBooth need?
Usually just a handful - often five to twenty good, varied photos of the subject. Quality and variety matter more than quantity: clean, well-lit images from several angles teach the subject better than dozens of similar ones.
Related terms
- LoRAA LoRA (Low-Rank Adaptation) is a small add-on file that teaches a base model a specific style, character, object or concept - without retraining the entire model.
- Textual inversionTextual inversion teaches an AI image model a new concept - a style, object or character - and packs it into a single reusable keyword called an embedding, without altering the model's own weights.
- CheckpointA checkpoint is a saved AI model file containing the full set of trained weights. It is the complete "brain" that generates images - swapping checkpoints changes the entire look and capability.
- ControlNetControlNet is an add-on that conditions a diffusion model on a reference structure - such as a pose skeleton, edge map or depth map - so you control composition precisely, not just with words.