Deepfake vs AI-generated: what each term means in detection benchmarks

This page belongs to the deepfake detection hub. It is the vocabulary note behind the four kinds of face deepfake page: that page tells the kinds apart by what each one changes, this one explains why "deepfake" and "AI-generated" are not interchangeable when a benchmark score is quoted.

01Two words, two different test sets

The two terms describe two different relationships between a fake and reality. A deepfake, in the sense every face benchmark uses, starts from a real recording and changes something in it: the identity of the face, the expression, the mouth movement, an attribute such as age. An AI-generated image starts from noise or from a text prompt and has no original underneath. The distinction matters for detection because the two leave different traces: a manipulated frame carries a boundary between altered and untouched pixels, while a generated frame is synthetic everywhere, background included.

Benchmarks follow that split in their construction. A deepfake dataset is built by taking source videos and running manipulation methods over them, so every fake has a matching real clip. A generated-media dataset is built by sampling generators, and the "real" half is a separate collection chosen to match in content. Each public set's build is recorded on the page listing which public sets hold which kind of fake.

02What counts as a deepfake in the benchmark papers?

The FaceForensics++ paper (ICCV 2019) set the working definition for face video. Its benchmark is "based on DeepFakes, Face2Face, FaceSwap and NeuralTextures", four automated methods applied to 1,000 original sequences drawn from 977 YouTube videos (arXiv 1901.08971). Two of the methods swap identity and two reenact expression, and none generates a face from nothing, which DF40's comparison table records as 2 face-swapping and 2 face-reenactment methods and no entire face synthesis.

The Celeb-DF paper (CVPR 2020) is narrower still. Its abstract opens with "AI-synthesized face-swapping videos, commonly known as DeepFakes", and the dataset contains 5,639 face-swapped videos of 59 subjects and nothing else (arXiv 1909.12962). The Google Deep Fake Detection set released two months before it is face swaps of consenting actors. When a paper from this period reports a deepfake detection score, the fakes behind it are manipulated real footage, and in most cases face swaps.

03What counts as AI-generated?

The generated-image literature uses a different vocabulary and different data. GenImage (2023) describes itself as a benchmark for "detecting AI-generated fake images" and holds 2,681,167 images, of which 1,331,167 are real and 1,350,000 are generated by eight generators across the 1,000 ImageNet classes (arXiv 2306.08571). The real images are ImageNet photographs of objects and animals, not faces, and every fake was produced from a label, with no source photograph manipulated. The generated-image benchmarks page covers that family.

For faces, the same category is what DF40 calls entire face synthesis, defined in its section 2.1 as a method that "generates a entirely synthesis face" with new identity, new attributes and new background at once. The generators are the GAN and diffusion families described on the fully generated faces page: VQGAN, StyleGAN2, StyleGAN3 and StyleGAN-XL on the GAN side, and DDPM, RDDM, PixArt-alpha, DiT-XL/2 and SiT-XL/2 on the diffusion side, by DF40's Table 2.

04How does DF40 split its 40 methods?

DF40 is the first face benchmark to carry both kinds in one set and to label them. Its Table 1 counts "10 FS methods, 13 FR methods, 12 EFS methods, and 5 FE methods": face swapping, face reenactment, entire face synthesis and face editing. The first two are deepfakes in the FaceForensics++ sense, the third is fully generated, and the fourth edits attributes of a real face. The same table counts earlier sets by the same scheme: DFDC has 8 methods (5 FS, 1 FR, 2 EFS) and ForgeryNet 15 (6 FS, 4 FR, 2 EFS, 3 FE), so the generated kind was present in those sets but in small numbers.

The distinction also shows in how the data were made. DF40's swapping and reenactment methods run over the real footage of FaceForensics++ and Celeb-DF, so each fake has a source clip. Its entire face synthesis methods are fine-tuned on that real data and then sampled, producing more than 48,000 images per method with no matching original, by the data scale column of Table 2. The DF40 dataset page describes the four protocols built on those splits.

Counts from the DF40 comparison table (arXiv 2406.13495, Table 1), read 10 Nov 2026; FaceForensics++ from its paper (arXiv 1901.08971).

05Does a detector trained on one kind catch the other?

DF40's Protocol 1 answers this directly. Xception trained on the face-swapping fakes of the FaceForensics++ domain scores a frame-level AUC of 0.991 on unseen face swaps, 0.892 on reenactments and 0.810 on entire face synthesis (Table 3). Trained on entire face synthesis instead, the same network scores 0.999 on synthesis, 0.807 on reenactment and 0.665 on face swaps. Each kind is easiest for a detector trained on it and hardest for a detector trained on the other.

The paper's reading is that the fake region matters: a model trained on fully generated faces "might be biased to 'expect' artifacts across the entire image, including the background", and so misses a swap that touched only the face. A score quoted as "deepfake detection accuracy" therefore says little about generated images, and a generated-image score says little about swaps, unless the test set is named.

06Why the terms get mixed up

Outside the papers, "deepfake" is used for any synthetic media, including fully generated faces and voices, and "AI-generated" is used for manipulated video. The definition of a deepfake page traces the word from its 2017 origin, where it meant face swaps, to its broader use. The benchmark usage has stayed narrow because the datasets fix it: a result on FaceForensics++ or Celeb-DF is a result on manipulated footage, and a result on GenImage is a result on generated images of objects. What a manipulated clip needs in order to pass as real is a separate question, answered on what makes a deepfake realistic.

The practical check is the test set, which is also the first of the five items a score has to state before it can be rerun, as the worked example on reproducing a detection result shows. If a reported figure comes from FaceForensics++, Celeb-DF, DFDC or the swapping and reenactment splits of DF40, it measures detection of manipulated real footage. If it comes from GenImage, the entire face synthesis split of DF40 or a StyleGAN face set, it measures detection of generated media. A page that quotes one kind of score as if it covered the other is comparing two experiments that share a word and nothing else.