Face swap, reenactment, attribute editing and full synthesis: the four kinds of face deepfake

This page is part of the deepfake detection hub and uses the vocabulary of the two datasets that define it: FaceForensics++, whose paper divides manipulation into "facial expression manipulation and facial identity manipulation", and DF40, which sorts its 40 techniques into face swap, face reenactment, entire face synthesis and face editing. The datasets table records which kinds each set contains.

The counts are DF40 Table 1 (arXiv 2406.13495, read 10 Nov 2026).

01Face swap

A swap takes a source identity and places it onto a target video, keeping the target's head, hair, body, lighting and movement. The FaceForensics++ paper describes two routes. FaceSwap is "a graphics-based approach to transfer the face region from a source video to a target video": landmarks are fitted, a 3D template is warped to the source face, and the result is blended in. DeepFakes, the method that gave the whole field its name, is learning-based: "two autoencoders with a shared encoder that are trained to reconstruct training images of the source and the target face", so that swapping the decoders renders the source face with the target's expression. Later swaps such as FaceShifter, added to FaceForensics++ in 2020, and the single method behind Celeb-DF improved the blending that earlier methods left visible.

The traces a swap leaves are at the boundary and in the identity. The new face has to be blended onto the old head, so colour, resolution and noise can change at the seam, and the swapped face has to be consistent from frame to frame while the head moves. Methods built on those cues, such as the blending-boundary detector Face X-ray, score well where the boundary exists and poorly where a generator leaves none: in the DeepfakeBench table Face X-ray reaches 0.9872 frame-level AUC on the Face2Face subset of FaceForensics++ and 0.5531 on DeeperForensics-1.0.

02Reenactment

Reenactment keeps the identity in the video and changes what the face does. The FaceForensics++ paper calls Face2Face "a facial reenactment system that transfers the expressions of a source video to a target video while maintaining the identity of the target person", a graphics method that fits a face model to both performers and re-renders the target's expressions. NeuralTextures is the learning-based counterpart in the same dataset; the authors use it to modify "the facial expressions corresponding to the mouth region", leaving the rest of the face as recorded. DF40 counts 13 reenactment methods among its 40, more than any other category, reflecting how the talking-head and lip-sync generators multiplied after 2019.

Because the identity is genuine, reenactment defeats any check that asks whether the face belongs to the person. Its traces are in the mouth and jaw, where the re-rendered region meets the original, and in the match between speech and lip movement when audio is involved.

03Attribute editing

Editing changes one property of a face and leaves the rest: age, hair colour, a smile, glasses. It is the smallest category in DF40, with 5 methods, and absent from FaceForensics++, Celeb-DF and the Deepfake Detection Challenge set, which is why detectors trained on those sets have rarely been tested on it. The manipulated region can be small and the identity is unchanged, so edits are the least studied and, in published tables, the least often scored kind.

04Entire face synthesis

Synthesis produces a face from nothing. The generator, adversarial or diffusion-based, is trained on a corpus of real faces and samples new ones; there is no target video, no boundary and no original to compare against. DF40 holds 12 such methods. The faces in the human studies on the can people spot AI faces page were made this way with StyleGAN2, and the AI-generated faces page covers how detectors look for them, mostly through traces of the generator's own processing rather than of any edit.

05Which datasets cover which kind

FaceForensics++ (2019) contains two swaps (DeepFakes, FaceSwap) and two reenactments (Face2Face, NeuralTextures), plus the FaceShifter swap added in 2020. Celeb-DF v2 (2020) is one improved swap. DeeperForensics-1.0 (2020) is one end-to-end swap with real-world perturbations. The Deepfake Detection Challenge set (2020) used eight methods described by its authors as "several Deepfake, GAN-based, and non-learned methods", all applied to paid actors. DF40 (2024) covers all four kinds with 40 techniques, and its authors' motivation is that FaceForensics++ "contains out-of-date forgery techniques from the past four years". The in-the-wild sets, WildDeepfake and Deepfake-Eval-2024, contain whatever circulated and do not label the kind. The counts and sources for each are on the datasets page, and the FaceForensics++ page describes its four methods in detail.

06Why detectors trained on one kind miss another

A detector learns whatever separates the fakes from the reals in its training set. On FaceForensics++ that is mostly swap and reenactment artefacts from four 2019 methods at known compression. A synthesised face has no boundary to find; an edited face has no identity change; a 2024 reenactment renders the mouth differently from NeuralTextures. The published numbers follow: in DeepfakeBench, ten of the fifteen detectors trained on FaceForensics++ c23 score above 0.93 on its own split, and all of them score between 0.55 and 0.95 on other sets, and DF40's authors report that current detectors "may not have obvious advantages over the baseline" when tested across its forgery types. The cross-dataset generalisation page gives the drops model by model.

07Liveness and injection as the delivery route

In an identity check the kind of fake matters less than how it reaches the camera. A swap or a reenactment can be shown to the camera on a screen, which is a presentation attack, or fed into the video stream in place of the camera, which is an injection attack. The liveness and injection attacks page covers how each is tested, and why the four kinds above all collapse into one question there: is a live person present?