How Deepfakes Are Made

Deepfakes use neural networks, but not one single architecture. Face-swap tools have often used autoencoders with shared representations; GANs can generate or refine synthetic faces; newer high-fidelity synthetic media also uses diffusion models. The common thread is that the system learns visual patterns well enough to produce convincing fake media.

The process may need only publicly available photos or video of the target, plus ordinary computing power. What once required a Hollywood studio can now be attempted with consumer tools, which is why deepfakes are no longer limited to state actors or big tech companies.