Mmmmmm some of the models on CivitAI, with the right workflows, could effectively create the most versatile "deep fakes" possible. You can put someone's face on a blank canvas and tell a model specifically trained for realistic pornography to paint a whole scene around the face. The only advantage here is that doing that is kinda pointless when you can just generate any random face to match your specifications. So ultimately much less harmful so long as the user isn't obsessing over representing a distinct living person.

Also, these are primarily still images. Some animation models exist, but that process is a lot more hit-and-miss. Overall though, I'd argue this whole use case is significantly less damaging than deep fakes due to principals.