Blog

Get Better DreamBooth Results Using Offset Noise

DreamBooth tends to overfit and produce weird artifacts, like extra heads or multiple faces, especially with only a few training images. One trick to.

Get Better DreamBooth Results Using Offset Noise

TheLastBen recently updated our fast stable diffusion template with the offset noise functionality in dreambooth!

To turn it on, start up the template and connect to the jupyter link. Then just flip the Offset_Noise variable to "True" and increase your total steps by 10-20% - that's it!

Jupyter notebook DreamBooth training settings with the Offset_Noise option circled in red

Apparently this can get you some really awesome outputs in the realm that SD usually can't touch ( very dark and very light images). For more on the research behind this, you can check out this research blog post

and this YouTube video: