Yes, and with less data than most people assume. Modern image-generation models can produce convincing renderings of a specific person from a handful of public photos. Some methods need only one or two images for a "personalization." The technical asymmetry is large, and it is not slowing down.
How face generation actually works
Modern image models are diffusion or transformer-based. They learn a statistical map from text descriptions to image patches across billions of training examples. When you prompt for a face, the model is sampling from a learned distribution conditioned on the prompt, not retrieving a specific image, but reconstructing a plausible one.
This is why "the model is just generating, not copying" is technically accurate but legally irrelevant. The output is novel; the underlying representation is learned from real faces, including yours if you have a public presence. The output can be statistically close to you because your face was a training signal.
Why a few photos is enough
Personalization techniques like LoRA, Dreambooth, or textual inversion can fine-tune a base model toward a specific identity from as few as 5 to 20 images. The base model already knows "human face." The fine-tune teaches it "this particular configuration of human face." Once tuned, the model can produce arbitrary new poses, expressions, and contexts, none of which existed in the training set.
The implication: if you have a public LinkedIn profile, a few news articles, and tagged social photos, you have already supplied enough signal for personalization. The defense is no longer information control (everyone has too much public exposure for that). The defense is rights control, making unauthorized commercial use legally costly.
What you cannot do
You cannot remove your face from already-trained models. "Machine unlearning" is an active research area, but production deployment is rare and incomplete. Once your face is in the weights, it is statistically present.
You also cannot rely on platform-level enforcement. Major model providers have content policies that prohibit non-consensual generation of real people, but enforcement is patchy and mostly reactive. The model itself is capable; the policy is voluntary.
What you can do
Reduce ongoing exposure: audit and clean up public photos that you do not need to be public. This will not retroactively help, but it slows future fine-tunes.
Use opt-out signals where they exist: some training datasets honor robots-style exclusions, opt-out registries, or image-level metadata. Compliance is voluntary, but compliant operators (the more risk-averse ones) tend to honor them.
License proactively: register your identity under defined terms, and make any commercial use of your face that is not on your terms clearly unauthorized. This converts open-ended exposure into a contract, and a contract is the basis for any successful enforcement action.
You cannot prevent the technical capability. You can change the legal calculus around its use.
Every AI-generated face is a legal risk. Unless it is licensed.
Protect your identity