Dreambooth finetuned models no longer producing good results in the latest releases

[Gman0909]

I have a few custom face models trained in Dreambooth, which used to produce excellent results, Over the past couple of days however something changed in the latest releases that has not only invalidated previous seeds (I can no longer replicate previous images under the same conditions, with hgypernetworks disabled and everything the same) but it's made it much more difficult to obtain good results even with the same exact prompts.

Specifically, it appears that any style queues (e.g. "by anna dittman") now almost completely override the likeness of the face, losing most of the similarity to the original. Keeping away from style cues in the prompt yields better likenesses, at the cost of stylistic predictability.

Has anyone experienced this? Is there some option that has been added that may be affecting my results?