Plain-English explanation
DreamBooth answers a specific question: how do you teach an image model a new, specific subject (not just "a dog" but your dog) using only a handful of photos, without the model forgetting everything else it knows or losing the ability to generate that subject in new poses, outfits, and settings? Its solution is to continue training the model directly on those few images, tying the subject to a rare or made-up word in the prompt (the original paper's examples used tokens like an uncommon identifier combined with the subject's class, e.g. "a photo of [x] dog"), while also training on a set of generic images of the same class to keep the model from forgetting what dogs in general look like.
How it works
DreamBooth fine-tunes the diffusion model's own weights (typically the image-generating network (UNet) and, optionally, the text encoder) using the small set of subject images paired with a prompt containing a distinctive placeholder token. A key technique the method introduces is prior preservation loss: alongside the few subject images, the training process also generates and trains on a batch of the model's own generic outputs for the subject's broader class (for example, generic dog images if the subject is a specific dog), which helps prevent the model's general understanding of "dog" from collapsing onto just the one training subject and helps it retain the ability to generate the subject in new contexts.
Because it updates real model weights rather than a small side file, DreamBooth training needs more GPU memory and time than LoRA or Textual Inversion training, and is also more prone to overfitting if hyperparameters (learning rate, training steps) aren't tuned carefully: the technique's own documentation describes it as "very sensitive to training hyperparameters." Many current tools run DreamBooth's training approach but save the result as a LoRA adapter instead of a full checkpoint (a hybrid sometimes called "DreamBooth-style LoRA training"), combining DreamBooth's training method with LoRA's lighter output format, Check which variant a given tool uses.
Adult-AI use
DreamBooth is the technique behind fully custom checkpoints built around one specific character, rather than a base model plus an add-on. Some creators use it (or DreamBooth-style training saved as a LoRA) to produce a checkpoint that reliably generates a particular original character in a very wide range of scenes and poses, more robustly than a lighter-weight adapter sometimes can. It's more resource-intensive to train than a LoRA, so it's less common for casual hobbyist use, but it shows up wherever someone wants maximum consistency and is willing to spend the extra GPU time and training data curation to get it.
How it differs from LoRA and Textual Inversion
DreamBooth, LoRA, and Textual Inversion all learn a new subject from a handful of images, but they store what they learn very differently:
- DreamBooth updates the model's actual weights through further training, generally producing a new full-size or merged checkpoint.
- LoRA freezes the base model and trains a small separate set of adapter weights instead, producing a lightweight file applied on top.
- Textual Inversion doesn't train the model's weights at all; it only learns a new text embedding tied to a placeholder word, leaving the model itself completely untouched.
DreamBooth is the heaviest of the three in training cost and file size, and the most likely to produce a subject that generalizes well across very different prompts, at the cost of longer training time and higher overfitting risk.
Common uses
- Training a highly consistent original character into a dedicated checkpoint for repeated, reliable use.
- Teaching a model a specific object, product, or art style that needs to generalize across many different prompts and contexts.
- Producing merge-ready checkpoints that other creators combine with other fine-tuned or base models.
Limitations and misconceptions
- Easy to overfit. Too many training steps or too few/too-similar source images can make the model reproduce training photos too literally, or "forget" how to generate the subject in new contexts.
- Not lightweight. Full DreamBooth training and output are far more resource-intensive than LoRA or Textual Inversion, both to train and (for full-checkpoint outputs) to store and distribute.
- Not the same as merging models. DreamBooth trains from example images; merging combines the weights of two already-trained checkpoints without new training data.
Privacy and safety considerations
Because DreamBooth can train a highly consistent likeness from a small number of reference photos, it carries the same consent risks as LoRA training but arguably more forcefully, since a DreamBooth-trained checkpoint can reproduce its subject with strong fidelity across a very wide range of scenes. Training this technique on a real, identifiable person's photos without their consent, particularly for sexual content, is harmful and increasingly illegal: legitimate use is limited to original characters or your own likeness with proper rights.