Dreambooth fine-tuning with prior preservation to reduce overfitting
When fine-tuning Stable Diffusion with Dreambooth, add a prior-preservation step that includes images from the target’s broader class (e.g., a generic “person” class when training a new face). During fine-tuning, these class images are intended to counterbalance the new concept’s tendency to overfit.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
The source describes Dreambooth as prone to quickly overfitting, especially for faces. Prior preservation is presented as a method to improve results and reduce overfitting by blending new-instance images with additional class examples during training.
What changed
Instead of fine-tuning only on the new subject’s images, training also incorporates automatically generated (or user-supplied) prior-preservation images representing the same class to be preserved.
Where it helps
- Training a new human face concept with Dreambooth
- Reducing overfitting when incorporating a new person identity
- Learning a new style or object concept where overfitting appears quickly
- Using Dreambooth on smaller face datasets where general class context helps
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
The source indicates that even with prior preservation, results may still show noise artifacts/blotches in some face-training cases. It also notes that the effect may be less significant for non-face subjects.
Sources
- Training Stable Diffusion with Dreambooth using Diffusers ↗
Hugging Face · primary · Published 2022-11-07 · Checked 2026-11-10
