This model is based on a concept from SDXL I use a lot, namely removing overtraining from the model to improve coherence and/or adherence. A side effect is it tends to strip away the safety tuning. It's a bit tedious and hard to target, and works totally differently in Flux. This was, unfortunately, and expensive model to train, since I had to generate the Negatives on the cloud using buzz... and I'm not entirely happy with the results.

This negative does strip away part of the safety tuning, but not as much as I would have hoped.