Stable Diffusion added a DreamBooth style training module over the Thanksgiving weekend and it is with great frustration that I'm here to share the results of the testing so far.
DreamBooth is a different style of training from the native Stable Diffusion training. It was originally cooked up by the trannies at Google and functions by removing everything from the starting database then replacing it with a database created solely from the trained images. This makes it
really good at learning the thing you give it. So far I'd say it stands head and shoulders above the results from the native Stable Diffusion training.
Sadly, we've now learned what happens when you merge the DreamBooth trained module with a different module. It gradually subsumes the library of the module you merged it with, which means everything is tainted with a touch of the merged results.
For instance, we trained on Tracy J Butler's style because we'd run it through the native SD training and wanted to compare. Well, everything -- no matter what style was requested -- took on a hint of Tracy J Butler's style. Super disappointing.
We did teach it Kiwis though, and I figured that would be cool to share since everyone here who's played with generating AI art probably knows how hard it is to get a decent kiwi bird.
