Fully AI-generated faces (GAN and diffusion): how detectors and datasets handle them

This page is part of the deepfake detection hub. It covers the kind of fake the kinds of face deepfake page calls entire face synthesis, and it is the technical companion to the can people spot AI faces page, where the same images defeated human observers.

01What an AI-generated face is, and how it differs from a face swap

A swap starts from footage of a real person and replaces one region; the rest of the frame is a genuine recording, and the manipulated region has to be joined to it. A generated face starts from noise. In an adversarial network the generator learns, against a discriminator that tries to tell its output from real photographs, to produce faces the discriminator cannot reject; Nightingale and Farid describe the process in their PNAS paper and used StyleGAN2, trained on 70,000 real faces, for their stimuli. In a diffusion model the network learns to remove noise step by step from a random field until a face remains. Either way the whole image is synthetic, so there is no seam, no identity mismatch and no original to compare against. Everything a detector can use is a property of the generator.

02GAN faces and the frequency cues detectors used

Two results from 2020 set the pattern. Wang, Wang, Zhang, Owens and Efros showed that "a standard image classifier trained on only one specific CNN generator (ProGAN) is able to generalize surprisingly well to unseen architectures, datasets, and training methods", across images from 11 CNN-based generators, "with careful pre- and post-processing and data augmentation" (arXiv 1912.11035). Generated images of that era shared something a classifier could learn once and reuse.

Frank, Eisenhofer, Schönherr, Fischer, Kolossa and Holz said what it was. "In frequency space, GAN-generated images exhibit severe artifacts that can be easily identified", and "these artifacts are caused by upsampling operations found in all current GAN architectures, indicating a structural and fundamental problem" (arXiv 2003.08685). The upsampling layers that enlarge a small latent image to full resolution leave regular patterns in the spectrum that no camera produces. For a few years that gave image-level detectors a cue that did not depend on the face at all.

03Diffusion faces and why earlier cues weaken

Diffusion models changed the processing path, and with it the traces. Corvi, Cozzolino, Zingarini, Poggi, Nagano and Verdoliva set out to "expose the forensics traces left by diffusion models" and to "study how current detectors, developed for GAN-generated images, perform on these new synthetic images, especially in challenging social-networks scenarios involving image compression and resizing" (arXiv 2211.00680). The framing of that study is the point for a reader of benchmark tables: a detector's published score was earned on a particular generator family, and a new family is a new test set, exactly as a new manipulation method is for face swaps. Compression and resizing, the ordinary life of an image on a platform, attack the frequency cues directly, since both operations rewrite the high-frequency content where the artefacts live. The DF40 dataset's authors make the same argument for video: FaceForensics++ "contains out-of-date forgery techniques from the past four years", and detectors honed on them cannot be assumed to generalise to current generators (arXiv 2406.13495).

04Datasets that cover generated faces

The video datasets that dominate detection results contain almost no synthesis. FaceForensics++, Celeb-DF, DeeperForensics-1.0 and the Deepfake Detection Challenge set are swaps and reenactments of real footage. DF40 (2024) is the first widely used set to cover the kind in depth, with 12 entire-face-synthesis methods among its 40, alongside 10 swaps, 13 reenactments and 5 editing methods. Deepfake-Eval-2024 includes 1,975 circulated images collected in 2024, generator unknown. The datasets page carries the counts; for generated stills specifically, the primary sources are the generator papers and the detection papers above, which released their own image sets rather than a shared benchmark.

05Published results on generated-face detection

This site's results page reproduces only results it can place in a table with train set, test set, protocol and metric, and the generated-image literature does not yet have the equivalent of DeepfakeBench's single shared evaluation; each paper tests on its own mix of generators. The figures that can be stated with their conditions are these. Deepfake-Eval-2024 reports that open-source image detectors' AUC fell by 45% on its circulated 2024 images compared with their original benchmarks. Wang et al. report generalisation from one GAN to ten other CNN-based generators under their augmentation regime. Both describe a moment, and the DF40 finding that current detectors "may not have obvious advantages over the baseline" across forgery types is the caution to carry.

06Can people tell?

No, not for good samples. Nightingale and Farid's participants classified StyleGAN2 faces and matched real photographs with 48.2% accuracy untrained and 59.0% after training with feedback, and rated the synthetic faces as more trustworthy. The can people spot AI faces page gives the study in full. The practical consequence is that the visible tells taught by games and guides belong to older generators and should not be relied on.

07Checking a photo yourself

Because the image is synthetic throughout, the useful checks are about the image's history rather than its pixels: where it first appeared, whether a reverse search finds the same face attached to a different name, whether the file carries content credentials, and whether the person it is said to show exists anywhere else. The checking a face video page orders those steps by evidence; the same order applies to a still, with the detector score last and its published limits in mind.